Este estudio presenta RareCode, un marco de aprendizaje profundo no supervisado que detecta anomalías en imágenes histopatológicas de cáncer colorrectal mediante el análisis de la Rareza de Activación del Código (CAR), con el potencial de ayudar en la preselección sin necesidad de datos de entrenamiento etiquetados.
Artículo de investigación
RareCode: un marco de aprendizaje profundo no supervisado para la detección de anomalías en imágenes histopatológicas de cáncer colorrectal
En este artículo
Resumen
Resumen
La detección no supervisada de anomalías en imágenes histopatológicas ha sido ampliamente explorada mediante enfoques basados en la reconstrucción que miden los errores de reconstrucción, sin embargo, estos métodos a menudo no logran captar variaciones patológicas sutiles a nivel semántico debido a la heterogeneidad inherente de las texturas del tejido. Para abordar esta limitación, este estudio presenta RareCode, un marco basado en la cuantificación vectorial que extiende el paradigma de detección más allá de la reconstrucción a nivel de píxeles, incorporando un análisis semántico de la activación del código. La innovación principal es la puntuación de Raridad en la Activación del Código (CAR), que caracteriza la frecuencia de activación de cada entrada del código durante el entrenamiento con muestras exclusivamente normales y marca como indicadores de anomalía las activaciones infrecuentes en la fase de inferencia, complementando así los errores de reconstrucción con discriminación a nivel semántico. Sobre esta base de mecanismo CAR de una sola escala, se introduce además un módulo de Código Jerárquico Multiescala (MHC), que emplea códigos de distintos tamaños con pesos de fusión aprendibles para capturar patrones patológicos que abarcan desde estructuras gruesas a nivel de tejido hasta detalles finos a nivel celular. Aprovechando este diseño multiescala, se generan mapas térmicos jerárquicos de anomalías en cada nivel de granularidad del código, proporcionando a los patólogos pistas visuales interpretables y multidimensionales que indican tanto la ubicación como el nivel estructural en el que ocurren las anomalías. Una validación cruzada de cinco pliegues en un conjunto de datos de imágenes histopatológicas de cáncer colorrectal con anotaciones clínicas obtenido del Hospital Popular de Shenzhen demuestra que RareCode alcanza un área bajo la curva (AUC) del 96,82 %, superando a los métodos básicos. El análisis por clases mostró un mejor desempeño en la detección de cáncer (AUC = 99,44 %, especificidad = 94,21 %) que en la detección de inflamación (AUC = 94,30 %, especificidad = 60,44 %). Estos hallazgos sugieren que el análisis CAR podría ofrecer un enfoque no supervisado prometedor para la detección de anomalías histopatológicas, con potencial para apoyar el precribado clínico en el diagnóstico del cáncer colorrectal.
Introducción
El cribado histopatológico confiable para el cáncer colorrectal sigue limitado por la anotación de expertos y la revisión manual, especialmente cuando las lesiones inflamatorias y las malignidades presentan morfologías superpuestas1,2. Estas limitaciones motivan enfoques computacionales no supervisados que aprenden a partir de tejido normal sin requerir anotaciones exhaustivas de anomalías3,4.
La urgencia clínica del cribado histopatológico automatizado queda subrayada por el creciente desequilibrio entre la demanda de diagnósticos y la experiencia patológica disponible, ya que la carga de trabajo diagnóstica por patólogo en Estados Unidos aumentó más del 40 % en una sola década mientras la plantilla de patólogos seguía disminuyendo5. En el cáncer colorrectal (CCR), los programas de cribado organizados se han asociado con reducciones de la mortalidad del 29-68 %6, aunque la capacidad de revisión patológica manual sigue limitada por la disponibilidad de personal. Un sistema automatizado de precribado que identifique de forma fiable los casos sospechosos para revisión prioritaria podría ayudar a aliviar esta carga y mejorar los tiempos de respuesta diagnóstica. Sin embargo, la implementación de tales sistemas requiere una alta sensibilidad para evitar omitir los verdaderos positivos y una especificidad adecuada para prevenir alarmas falsas excesivas7.
La detección no supervisada de anomalías (UAD) se ha vuelto cada vez más relevante en el análisis de imágenes médicas8 porque puede modelar distribuciones de tejidos normales sin necesidad de etiquetas de entrenamiento anormales3,9,10. Los modelos basados en reconstrucción, incluidos los autoencoders, los autoencoders variacionales, las redes generativas adversarias y las variantes aumentadas con memoria, identifican anomalías mediante errores de reconstrucción, pero los decodificadores de alta capacidad aún pueden reconstruir regiones anómalas con alta fidelidad11,12,13,14,15,16,17. Los métodos basados en características, como PatchCore y PaDiM, utilizan representaciones previamente entrenadas, aunque las características aprendidas a partir de imágenes naturales pueden no capturar completamente las atipias microscópicas en histopatología18,19,20,21. Los modelos fundamentales específicos de la patología y los detectores de anomalías basados en difusión ofrecen alternativas poderosas, pero sus requisitos de datos o costos computacionales pueden limitar su uso directo en cribado de alto rendimiento22,23. Métodos recientes automatizados y evolutivos, como EvoAAE24 y MoARNN-AM25, ilustran aún más el valor de la optimización adaptativa del modelo para la detección de anomalías, aunque sus contextos de aplicación difieren del análisis de imágenes histopatológicas. En contraste, los métodos basados en cuantificación vectorial (VQ) imponen restricciones de diccionario discreto que pueden reducir la correspondencia de identidad; sin embargo, los enfoques actuales basados en VQ aún dependen principalmente de errores de reconstrucción espacial y subutilizan la información semántica contenida en los patrones de activación del diccionario26,27.
Aunque los métodos mencionados anteriormente demuestran un rendimiento prometedor en dominios generales, su aplicación a escenarios histopatológicos complejos enfrenta desafíos específicos. Las imágenes histopatológicas se caracterizan por una heterogeneidad tisular compleja a través de múltiples niveles estructurales28. En la práctica, el análisis de imágenes suele operar sobre fragmentos locales extraídos de secciones de tejido, y el diagnóstico patológico es inherentemente multiscale: los patólogos evalúan la morfología glandular y tisular a baja magnificación, mientras examinan el pleomorfismo nuclear y las figuras mitóticas a alta magnificación29. Identificamos tres limitaciones clave de los enfoques actuales: Primero, las anomalías patológicas y los tejidos normales presentan una gran similitud en características visuales de bajo nivel, como los estilos de tinción y las texturas locales, lo que provoca que los métodos basados en reconstrucción fallen al detectar anomalías sutiles, ya que tanto muestras normales como anómalas pueden producir una calidad de reconstrucción similar a nivel semántico. Segundo, la mayoría de los métodos existentes emplean extracción de características a una única escala, lo que dificulta captar simultáneamente características anómalas a nivel tisular y celular30,31. Tercero, aunque los métodos predominantes pueden generar mapas de calor a nivel de píxel, a menudo no logran capturar de forma intuitiva los atributos jerárquicos de las anomalías, lo que limita la utilidad de dichos modelos como herramientas de ayuda diagnóstica clínica.
Para abordar los desafíos anteriores, proponemos RareCode, un marco de trabajo para detección de anomalías sin supervisión (UAD) que explora patrones de activación de código y representaciones jerárquicas de características en múltiples niveles de abstracción. El marco introduce tres componentes principales: (1) Mecanismo de Puntuación de Raridad en la Activación del Códig (CAR), que calcula puntuaciones de rareza basadas en las frecuencias de activación de entradas entrenadas exclusivamente con muestras normales, distinguiendo muestras normales de anómalas a nivel semántico y proporcionando señales discriminativas complementarias a los errores tradicionales de reconstrucción espacial. (2) Arquitectura de Fusión de Códigos Jerárquicos Multiescala (MHC), que captura características patológicas en múltiples niveles de granularidad—desde patrones estructurales gruesos hasta detalles celulares finos—mediante el uso de códigos con capacidades variables, logrando una fusión adaptativa a través de pesos aprendibles. (3) Módulo Jerárquico de Localización Interpretable, que aprovecha la arquitectura multiescala para generar mapas de calor de anomalías con diferentes niveles de granularidad, proporcionando a los patólogos referencias diagnósticas semánticamente interpretables a múltiples escalas.
La hipótesis principal de este estudio fue que los perfiles de frecuencia de activación del código derivados de códigos entrenados exclusivamente con imágenes histopatológicas normales del colon y recto codificarían información semántica relevante para anomalías más allá del error de reconstrucción a nivel de píxeles, y que la integración multiescala de estas señales complementarias mejoraría la discriminación de imágenes que contienen tejido canceroso o inflamatorio frente a imágenes normales, en comparación con métodos representativos de DIA, evaluado principalmente mediante el AUC. Para probar esta hipótesis, evaluamos RareCode mediante validación cruzada de 5 grupos en un conjunto de datos de CCR anotado clínicamente y realizamos análisis de ablación y de localización para examinar las contribuciones y la interpretabilidad de sus componentes principales.
Protocolo
This study was approved by the Clinical Research Ethics Committee of Shenzhen People's Hospital (Approval No. LL-KY-2025300-01). The requirement for informed consent was waived by the ethics committee because this retrospective study used existing histopathological images and clinical materials without direct patient contact or intervention. All clinical data and histopathological images were de-identified before analysis, and no personally identifiable information was used in this study. All data used in this research were handled in accordance with the ethical standards of the institutional and national research committees.
Framework overview
The overall architecture of the proposed UAD framework, RareCode, is illustrated in Figure 1. The framework employs a parallel dual-branch architecture. For each 512 x 512 input image, non-overlapping 32 x 32 and 64 x 64 patches are extracted as fine- and coarse-scale inputs, respectively. The 64 x 64 patches are resized to 32 x 32 before being passed to the network so that both branches share the same encoder-decoder input size while preserving different receptive fields. Encoders in each branch map the extracted features into a latent space, where the MHC module imposes discretization constraints. Subsequently, decoders reconstruct the image from quantized features to compute spatial-domain reconstruction errors. The CAR mechanism then measures the degree of semantic-level anomaly by profiling codebook activation frequencies. Finally, the model fuses reconstruction scores and CAR scores through a weighted combination to obtain image-level anomaly scores. The RareCode framework is trained end-to-end using only normal samples, requiring no anomaly annotations. The dual-branch design was used to capture both local cellular features and broader glandular structures. The MHC module was used to model these features with codebooks of different capacities.
Encoder-decoder architecture
RareCode constructs structurally independent encoders and decoders for the fine- and coarse-scale branches. Each branch uses the same encoder-decoder design but does not share parameters. The encoder consists of four convolutional blocks, each including a 3 x 3 convolution, batch normalization, ReLU activation, and dropout. The channel width increases from 64 to 128, then to 256, and the final feature map is flattened and projected into a 64-dimensional latent embedding. The decoder mirrors the encoder with a fully connected projection layer followed by four transposed convolution layers, reconstructing each patch to the original input size of the network. Reconstruction error is calculated as the mean squared error between the input and reconstructed patches. By combining this encoder-decoder structure with discrete codebook constraints, RareCode learns compact representations of normal tissue and measures deviations during inference.
Multi-scale hierarchical codebook
To capture pathological features at multiple levels of abstraction—from broad tissue patterns to localized cellular variations—the MHC module employs K discrete codebooks 
with varying capacities n1, n2, ..., nK. In the final FourScales configuration, each branch contains four codebooks with 64, 128, 256, and 512 entries, respectively, and each codebook entry has a 64-dimensional embedding. For each encoded patch feature, the nearest codebook entry is selected according to Euclidean distance. The quantized outputs from different codebooks are then fused using learnable weights normalized across codebooks. Smaller codebooks, due to stronger compression constraints, are expected to encode coarse-grained prototypical patterns that abstract away local variations, whereas larger codebooks preserve finer-grained features that capture more specific structural characteristics. For each codebook 
, where ek(i) 
Rd denotes the i-th embedding vector in the k-th codebook, continuous features are mapped to the discrete codebook space via nearest neighbor lookup (Eq. 1 and 2):
To achieve adaptive fusion of multi-scale features, we introduce learnable weights . After softmax normalization, weighted summation is performed over the quantized outputs from each codebook (Eq. 3):
where σ(·) denotes the softmax function, ensuring that weights satisfy Σk σ(wk) = 1. This design enables the model to automatically adjust the contribution ratios of codebooks at different granularities based on the semantic content of the input features.
To optimize codebook learning, we adopt the standard VQ loss function (Eq. 4):
where sg[·] denotes the stop-gradient operation, and β = 0.25 is the commitment loss weight coefficient. The first term encourages codebook vectors to move toward encoder outputs, while the second term encourages encoder outputs to remain consistent with their corresponding codebook vectors. The complete mechanism of multi-scale VQ is illustrated in Figure 2.
Codebook activation rarity scoring
CAR measures semantic-level anomalies based on codebook activation statistics estimated from normal training samples. After model training, all normal training images were passed through the encoder and MHC module without data augmentation. For each branch and each codebook, the assignment index of every patch feature was recorded, and the number of assignments to each codebook entry was counted. The counts were then normalized to obtain the activation-frequency distribution for that codebook (Eq. 5):
During inference, each test image was processed through the same patch extraction, encoding, and nearest-neighbor codebook assignment procedure. For each assigned codebook entry, the rarity score is calculated as the negative logarithm of its activation frequency in the normal training set (Eq. 6):
where ε is a small constant used for numerical stability. A low-frequency codebook entry, therefore, receives a higher rarity score, suggesting that the corresponding patch feature is less consistent with the learned normal-tissue distribution. Patch-level rarity scores were averaged within each image and across the fine- and coarse-scale branches to obtain an image-level rarity response.
To complement activation-frequency rarity, we also compute a percentile-based quantization distance score (Eq. 7):
This score is derived from the Euclidean distance between each encoded feature vector and its nearest codebook entry. The distance distribution is estimated from normal training samples, and test-sample distances are converted into percentile scores. Larger percentile scores indicate that the encoded feature is more difficult to represent using the learned normal codebook prototypes.
The final CAR score combines activation rarity and quantization distance across all codebooks using the learnable codebook weights (Eq. 8):
The same normalized codebook weights used for multi-scale feature fusion are applied to score-level fusion, maintaining consistency between representation learning and anomaly scoring. The resulting CAR score is calculated at the image level and is then combined with the reconstruction score to obtain the final anomaly score. The overall CAR scoring mechanism is illustrated in Figure 3.
Final anomaly score computation
We fuse spatial-domain reconstruction errors with semantic-level CAR scores. Before fusion, both scores are separately normalized using percentile-based min-max normalization to reduce the influence of extreme outliers. The final image-level anomaly score is then defined as (Eq. 9):
where Srecon denotes the normalized mean squared reconstruction error, reflecting sample reconstruction quality in pixel space; SCAR is the CAR score described above, capturing degrees of semantic-level anomaly. The hyperparameter α [0,1] controls the relative weighting between the two components, with optimal values determined on validation sets.
Training objective
The RareCode framework is trained end-to-end using only normal samples. The total training loss comprises the following components (Eq. 10):
Each term is defined as follows: Reconstruction Loss: 
, measuring the mean squared error between input patches p and reconstructed patches D(zq) after quantization. This loss is applied separately to the 32 × 32 patch branch (
) and 64 × 64 patch branch (
) to ensure effective learning of features at both spatial scales.
Dataset description
To evaluate RareCode, experiments were conducted on a clinically annotated CRC histopathological image dataset comprising images of colorectal tissue sections stained with hematoxylin and eosin (H&E), obtained from Shenzhen People's Hospital. All images in this dataset originate from real clinical cases. Images were digitized at 40x magnification with an original resolution of 1024 × 1024 pixels. To ensure annotation reliability and quality, all sample category labels were jointly reviewed and confirmed by two or more senior professional pathologists following a double-blind protocol.
Based on histopathological characteristics, the dataset was categorized into three classes: normal tissue (1,226 images), cancer (1,215 images), and inflammation (1,214 images). In the binary UAD setting, cancer and inflammation samples were treated as anomalous, while normal samples served as the reference distribution. This formulation reflects the intended triage task of separating cases requiring further pathological review from those that are morphologically normal.
All H&E-stained images were resized to 512 × 512 pixels before being used as network input. Stratified 5-fold cross-validation was used for evaluation. In each fold, the normal samples were first divided into training, validation, and held-out test subsets. RareCode was trained only on the training normal samples. The validation subset, which also contained only normal samples, was used for model selection, hyperparameter tuning, and selection of the fusion weight α. The held-out test set, consisting of unseen normal samples and anomalous samples including cancer and inflammation, was used only for final performance evaluation and was not used for model training, hyperparameter selection, or α selection. The representative categories, dataset structure, and validation protocol are shown in Figure 4.
Implementation details
The proposed RareCode framework and comparison baselines were implemented using PyTorch. All experiments were conducted on a workstation equipped with a single NVIDIA GeForce RTX 4060 Ti GPU (16 GB). For network architecture configuration, to capture features at different spatial extents, the dual-branch architecture processes non-overlapping patches at two scales: 32 × 32 pixels (smaller receptive field, capturing local texture patterns) and 64 × 64 pixels (larger receptive field, capturing broader spatial context), respectively. The continuous feature embedding dimension produced by both branch encoders is uniformly set to 64. In the MHC module, we configure four codebooks with different capacities for each branch, with discrete prototype quantities of 64, 128, 256, and 512, respectively.
During optimization, the network was trained end-to-end for 50 epochs with a batch size of 8. We employ the AdamW optimizer for weight updates, with an initial learning rate of 5 × and weight decay of to prevent overfitting. Additionally, to ensure smooth training convergence, a cosine annealing learning rate schedule was used, enabling finer optimization in later training stages. Patch sizes and codebook sizes were selected to balance multi-scale representation and computational cost. The final FourScales configuration was supported by ablation analysis, and α was selected on the validation set before test-set evaluation.
For the decoder-complexity ablation, a complex-decoder variant was implemented based on the FourScales configuration. This variant used the same data split, codebook sizes, training epochs, optimizer, learning rate, validation-based α selection, and evaluation metrics as the FourScales model. The basic decoder was replaced with an EMCAD-style decoder containing multi-scale depthwise convolution blocks and Convolutional Block Attention Modules (CBAM). After each of the first three transposed-convolution upsampling stages, parallel 3 × 3, 5 × 5, and 7 × 7 depthwise convolutions were applied, followed by 1 × 1 pointwise fusion, batch normalization, ReLU activation, a residual connection, and CBAM attention. CBAM included channel attention based on average- and max-pooling descriptors and spatial attention using a 7 × 7 convolution. The final stage used transposed convolution and sigmoid activation to reconstruct 32 × 32 patches.
Resultados
Comparison with baseline methods
To evaluate the RareCode framework, multiple representative UAD methods were selected as baselines, spanning reconstruction-based methods (CAE32, VAE12, SAE33, MAE34, PatchSAE35), memory-augmented reconstruction (MemAE36), knowledge distillation (STFPM37), synthetic anomaly generation (DRAEM38), and pretrained feature extraction (PatchCore18). Baseline selection focused on methods that can be trained or applied under comparable computational budgets and data assumptions (access only to unlabeled normal training samples), thereby ensuring that performance differences reflect methodological contributions rather than disparities in pretraining data scale. To ensure fair comparison, all methods were evaluated using the same 5-fold data splits, preprocessing pipeline, evaluation metrics, and computational environment. Models were trained under the same UAD protocol using only normal training samples, with hyperparameters and thresholds selected on the validation set and final performance assessed only on the held-out test set. Table 1 summarizes 5-fold cross-validation results, and Figure 5 provides multi-metric radar chart comparisons.
As shown in Table 1 and Figure 5, RareCode achieves favorable detection performance and statistical stability on the CRC dataset. In terms of AUC, RareCode (96.82 ± 0.23%) outperforms multiple reconstruction-based baseline models, including MemAE (94.97 ± 0.81%). Statistical analysis confirms that RareCode outperforms all reconstruction-based baselines (paired t-tests, p < 0.05), with the improvement over MemAE reaching statistical significance (p < 0.01). RareCode exhibits low performance variance (standard deviation of 0.23%), indicating consistent cross-fold stability. The limited effectiveness of DRAEM, based on synthetic anomalies, likely reflects the fact that simple texture overlay strategies cannot adequately simulate complex pathological atypia.
Regarding specificity, RareCode achieves 58.24 ± 5.99%, outperforming all comparison methods in reducing false positive rates. This represents approximately a 25-percentage-point improvement over the reconstruction-only baseline (NoCAR, 33.76%), demonstrating the CAR mechanism's contribution to reducing false positives. The lower specificity of STFPM and PatchCore, which rely natural-image-pretrained features, may partially reflect the domain gap between natural and histopathological images. Notably, STFPM and DRAEM exhibit high specificity variance (±33.53% and ±7.09%), with per-fold specificity ranging from near-zero to moderate levels, raising concerns about their deployment reliability.
Per-class detection performance
Per-class analysis was conducted by separately comparing cancer and inflammation samples against normal samples (Table 2). RareCode achieved higher performance for cancer detection (AUC = 99.44 ± 0.12%, recall = 98.13 ± 0.29%, specificity = 94.21 ± 2.22%) than for inflammation detection (AUC = 94.30 ± 0.44%, recall = 96.55 ± 0.78%, specificity = 60.44 ± 4.34%). These findings suggest that RareCode may show stronger discrimination for malignant abnormalities than for inflammatory changes, whereas the inflammation-normal boundary appears to contribute substantially to the reduced overall specificity.
Ablation study
To investigate the contributions of key components within the RareCode framework, ablation experiments were conducted to assess the impact of the CAR mechanism and the MHC module on detection performance. Results are detailed in Table 3. As shown in Figure 6A, adding CAR to the reconstruction-only baseline improved AUC from 92.81% to 95.92%, and multi-scale codebook fusion further increased AUC to 96.82%. Figure 6B shows that CAR also improved specificity, increasing it from 33.76% in the NoCAR baseline to 58.24% in the full FourScales model. These results support the complementary value of codebook activation rarity and multi-scale fusion.
Additionally, a decoder-complexity ablation was conducted using the FourScales configuration. As shown in Table 3, the complex-decoder variant with multi-scale depthwise convolution blocks and Convolutional Block Attention Modules (CBAM) showed lower AUC than the basic FourScales model (95.17 ± 0.44% vs. 96.82 ± 0.23%). This result suggests that, within the present VQ-AE-based RareCode setting, increasing decoder capacity did not improve anomaly detection performance and may reduce the effectiveness of the codebook-constrained representation.
Hierarchical spatial response analysis
To examine the spatial responses generated by RareCode, patch-level reconstruction error and codebook rarity maps were aggregated at the image level and compared between normal and anomalous samples. Energy ratio, Cohen's d, and AUC were used to quantify image-level response separation. Detailed results are presented in Table 4 and Figure 7 and Figure 8.
Results show that codebook rarity scores at all scales exhibit response elevation on anomalous samples (energy ratios > 1). Smaller-capacity codebooks (Codebook 64) achieve stronger localization-level discrimination (78.79% AUC), consistent with the expectation that higher compression rates encourage learning of more abstract prototypical patterns that are more sensitive to deviations from normal tissue archetypes. Reconstruction error alone provides substantial anomaly capture capability (93.31% AUC), while CAR scores offer complementary signals from the semantic frequency dimension.
Qualitative examination of Figure 7 showed that elevated responses in cancer samples overlapped with irregular and crowded glands, nuclear enlargement, hyperchromasia, pseudostratification, and loss of epithelial polarity. In inflammatory samples, stronger responses were observed around distorted crypts and dense inflammatory-cell infiltrates. Lower-capacity codebooks tended to capture broader architectural deviations, whereas higher-capacity codebooks produced more localized cellular-level responses. These findings provide qualitative morphological interpretation but not pixel-level validation.
Fusion parameter analysis
The fusion weight α balances the contributions of spatial reconstruction error and the CAR score to the final anomaly score. Optimal α values for each fold were determined via a grid search (step size 0.05) on the validation sets, with the results presented in Table 5 and Figure 9. Optimal α values primarily concentrate in the 0.85-0.95 range, with a mean of 0.90 ± 0.05. Under optimal configurations, the model achieves an average AUC of 96.82 ± 0.23% on test sets. Higher optimal weights (approximately 0.90) suggest that CAR scores contribute more to final detection than reconstruction error, while reconstruction error provides auxiliary local constraints. Compared with the α = 0 setting in the sensitivity analysis (AUC = 93.29%), the validation-selected fusion setting improved AUC by approximately 3.53 percentage points.
To summarize the present work presents the RareCode framework for unsupervised anomaly detection in histopathological images, aiming to mitigate the identity-mapping problem encountered by traditional generative models. The CAR mechanism reduces over-reliance on pixel-level reconstruction errors, while the MHC module provides hierarchical feature representations at multiple granularities, enabling complementary anomaly discrimination at different levels of abstraction. Experiments on the CRC dataset demonstrate that RareCode achieves 96.82% AUC, with per-class analysis showing effective cancer detection (AUC = 99.44%) and stronger discrimination for cancer than inflammation, suggesting that inflammation-normal overlap remains a major challenge. Ablation studies confirm that integrating discrete semantic features from VQ with multi-scale tissue morphological representations improves detection performance. RareCode's high recall (98.40%) and moderate specificity (58.24%) suggest its potential as a pre-screening tool for prioritizing suspicious histopathological images; however, its actual effect on pathologist workload will require prospective workflow evaluation.
Data Availability:
The CRC histopathological image dataset used in this study was obtained from Shenzhen People's Hospital under institutional ethics approval (Approval No. LL-KY-2025300-01). Due to patient privacy and institutional data sharing policies, the dataset is not publicly available. Access may be granted upon reasonable request to the corresponding author, subject to institutional approval and data use agreements. The source code for the RareCode framework is publicly available at https://github.com/XL-alg/RareCode.
Figure 1: Overall architecture and data flow of the RareCode framework. A 512 × 512 H&E-stained histopathological image is divided into non-overlapping 32 × 32 and 64 × 64 patches as fine- and coarse-scale inputs. The two branches encode patches into latent embeddings, apply Multi-scale Hierarchical Codebook (MHC) discretization, and reconstruct patches to compute reconstruction-error scores. In parallel, the Codebook Activation Rarity (CAR) mechanism estimates semantic rarity from codebook activation-frequency profiles learned from normal training samples. Normalized reconstruction and CAR scores are then fused to generate the final image-level anomaly score, while patch-level responses are projected back to the image plane for hierarchical localization. Please click here to view a larger version of this figure.
Figure 2: Multi-scale vector quantization mechanism. (A) Distance computation between encoder output features and codebook embedding vectors. (B) Nearest neighbor selection and weighted fusion across codebooks with capacities of 64, 128, 256, and 512. (C) Reconstruction from quantized features via transposed convolutional decoder. Please click here to view a larger version of this figure.
Figure 3: CAR scoring mechanism. This figure illustrates four stages: Stage 1: activation frequency statistics are computed solely on normal training samples. Stage 2: test samples obtain activation indices and distances through encoder and vector quantization. Stage 3: rarity scores and distance scores are computed based on training-set frequency profiles. Stage 4: scores from each codebook scale are fused through learnable weights to produce the final CAR score. Please click here to view a larger version of this figure.
Figure 4: CRC dataset composition and experimental protocol. (A–C) Representative H&E-stained histopathological images of normal colorectal tissue, colorectal cancer, and colorectal inflammation. (D) 5-fold cross-validation protocol for UAD, in which normal samples were used for training and validation, while held-out normal and anomalous samples were used for testing. Images were digitized at 40x magnification and processed into 32 x 32 and 64 x 64 patches. Green, light green, and orange indicate training, validation, and test sets, respectively. Please click here to view a larger version of this figure.
Figure 5: Multi-metric comparison radar chart of anomaly detection methods. Radar chart comparing RareCode against baseline methods across AUC, average precision (AP), F1-score, precision, recall, and precision at 90% recall (P@R90). All values represent mean performance across 5-fold cross-validation. Please click here to view a larger version of this figure.
Figure 6: Incremental contribution analysis of ablation study components. (A) AUC performance comparison showing the progressive improvement from reconstruction-only baseline (NoCAR) through single-codebook to multi-scale configurations. (B) Specificity performance comparison demonstrating the CAR mechanism's effect on false positive reduction. All error bars represent standard deviation (SD) across 5-fold cross-validation. Please click here to view a larger version of this figure.
Figure 7: Multi-scale codebook hierarchical localization heatmaps. Representative examples are shown for (A) normal, (B) cancer, and (C) inflammation samples. Each row includes the original image, overlay, reconstruction-error map, rarity-score maps from codebooks of different capacities (64, 128, 256, and 512), and the final fused localization map. Smaller-capacity codebooks highlight coarse-grained patterns through stronger compression abstraction, whereas larger-capacity codebooks preserve finer structural details. High-response regions qualitatively correspond to cancer- or inflammation-related histopathological features, but were not validated with pixel-level expert annotations. Please click here to view a larger version of this figure.
Figure 8: Distribution histograms of different score types. Histograms compare score distributions between normal and anomalous samples for (A) reconstruction error, (B) combined CAR score, (C) Codebook 64, (D) Codebook 128, (E) Codebook 256, and (F) Codebook 512. Green and red histograms indicate normal and anomalous samples, respectively. Please click here to view a larger version of this figure.
Figure 9: Alpha parameter sensitivity analysis. (A) Validation-set AUC used for α selection. (B) Test-set AUC under different α values. The validation-selected α values ranged from 0.85 to 0.95, with a mean of 0.90 ± 0.05, consistent with Table 5. AUC variation remained below 0.5% when α varied within the [0.70, 1.00] range, indicating stable performance across a wide parameter range. Please click here to view a larger version of this figure.
| Method | AUC (%) | AP (%) | F1 (%) | Precision (%) | Recall (%) | Specificity (%) | P@R90 (%) |
| CAE | 90.37 ± 0.94 | 98.62 ± 0.18 | 95.34 ± 0.15 | 91.64 ± 0.40 | 99.35 ± 0.22 | 28.14 ± 3.88 | 95.66 ± 0.32 |
| SAE | 90.58 ± 0.94 | 98.66 ± 0.18 | 95.34 ± 0.15 | 91.67 ± 0.40 | 99.32 ± 0.24 | 28.46 ± 3.97 | 95.71 ± 0.30 |
| VAE | 93.60 ± 0.82 | 99.13 ± 0.12 | 95.86 ± 0.19 | 92.81 ± 0.49 | 99.12 ± 0.29 | 39.07 ± 4.70 | 96.94 ± 0.43 |
| MAE | 91.65 ± 2.75 | 98.76 ± 0.50 | 95.61 ± 0.60 | 92.87 ± 1.56 | 98.55 ± 0.55 | 39.74 ± 14.32 | 96.55 ± 0.97 |
| MemAE | 94.97 ± 0.81 | 99.35 ± 0.11 | 95.78 ± 0.33 | 92.82 ± 1.07 | 98.95 ± 0.60 | 39.15 ± 9.99 | 97.57 ± 0.33 |
| PatchSAE | 94.86 ± 0.36 | 99.34 ± 0.05 | 95.39 ± 0.13 | 92.32 ± 0.27 | 98.67 ± 0.12 | 34.91 ± 2.57 | 98.13 ± 0.24 |
| STFPM | 90.23 ± 3.05 | 98.70 ± 0.44 | 94.32 ± 0.21 | 91.90 ± 3.51 | 97.14 ± 3.38 | 29.82 ± 33.53 | 95.93 ± 1.96 |
| DRAEM | 76.34 ± 2.82 | 95.34 ± 0.90 | 94.19 ± 0.07 | 89.39 ± 0.65 | 99.56 ± 0.65 | 6.19 ± 7.09 | 92.69 ± 0.57 |
| PatchCore | 74.64 ± 1.82 | 94.76 ± 0.55 | 94.73 ± 0.05 | 90.52 ± 0.15 | 99.36 ± 0.12 | 17.49 ± 1.54 | 92.54 ± 0.15 |
| RareCode | 96.82 ± 0.23 | 99.59 ± 0.03 | 96.63 ± 0.15 | 94.93 ± 0.67 | 98.40 ± 0.48 | 58.24 ± 5.99 | 98.80 ± 0.10 |
Table 1: Comparison results with baseline methods (5-fold cross-validation). Detection performance of RareCode and nine baseline UAD methods on the CRC dataset. Metrics include AUC, average precision (AP), recall, specificity, F1-score, precision at 90% recall (P@R90), and precision. All values represent mean ± SD across 5-fold cross-validation. Bold values indicate the best performance for each metric.
| Category | AUC (%) | AP (%) | F1 (%) | Recall (%) | Specificity (%) |
| Cancer | 99.44 ± 0.12 | 99.86 ± 0.03 | 98.34 ± 0.17 | 98.13 ± 0.29 | 94.21 ± 2.22 |
| Inflammation | 94.30 ± 0.44 | 98.48 ± 0.12 | 93.44 ± 0.29 | 96.55 ± 0.78 | 60.44 ± 4.34 |
Table 2: Per-class detection performance across anomaly subtypes. Separate evaluation of RareCode's detection performance for cancer and inflammation samples against normal samples. Per-class metrics, including AUC, AP, F1-score, recall, and specificity, were computed using class-specific optimal thresholds.
| Variant | Codebook Config | CAR | AUC (%) | AP (%) | F1 (%) | Specificity (%) | P@R90 (%) |
| NoCAR |  | | 92.81 ± 0.12 | 99.05 ± 0.02 | 95.30 ± 0.08 | 33.76 ± 3.82 | 96.72 ± 0.12 |
| SingleCodebook | [256] |  | 95.92 ± 0.40 | 99.47 ± 0.05 | 96.25 ± 0.14 | 55.37 ± 6.51 | 98.31 ± 0.43 |
| TwoScales | [128, 512] | | 96.57 ± 0.27 | 99.56 ± 0.04 | 96.45 ± 0.12 | 57.75 ± 4.14 | 98.75 ± 0.12 |
| ThreeScales | [128, 256, 512] | | 96.36 ± 0.37 | 99.53 ± 0.05 | 96.52 ± 0.17 | 57.99 ± 4.65 | 98.60 ± 0.23 |
| FourScales | [64, 128, 256, 512] | | 96.82 ± 0.23 | 99.59 ± 0.03 | 96.63 ± 0.15 | 58.24 ± 5.99 | 98.80 ± 0.10 |
| FourScales + complex decoder | [64, 128, 256, 512] | | 95.17 ± 0.44 | 99.39 ± 0.06 | 95.32 ± 0.20 | 53.83 ± 21.45 | 98.40 ± 0.37 |
Table 3: Ablation study results. Performance comparison of RareCode configurations with progressive addition of components: reconstruction-only baseline (NoCAR), single-codebook CAR (SingleCodebook), and multi-scale configurations (TwoScales, ThreeScales, FourScales). An additional decoder-complexity ablation using the FourScales configuration is also included. Metrics include AUC, average precision (AP), F1-score, specificity, and precision at 90% recall (P@R90).
| Score Type | Energy Ratio | Cohen's d | AUC (%) |
| Reconstruction Error | 2.623 | 2.006 | 93.31 |
| Combined CAR | 1.078 | 1.002 | 74.85 |
| Codebook 64 | 1.109 | 1.207 | 78.79 |
| Codebook 128 | 1.085 | 1.067 | 76.19 |
| Codebook 256 | 1.065 | 0.916 | 72.80 |
| Codebook 512 | 1.064 | 0.833 | 70.38 |
Table 4: Image-level analysis of localization-derived responses. Image-averaged reconstruction error and codebook rarity responses were compared between normal and anomalous samples using the energy ratio, Cohen's d, and AUC.
| Fold | Optimal α | Val AUC (%) | Test AUC (%) |
| 1 | 0.95 | 96.22 | 96.91 |
| 2 | 0.9 | 96.38 | 96.39 |
| 3 | 0.85 | 96.71 | 96.82 |
| 4 | 0.85 | 96.22 | 96.93 |
| 5 | 0.95 | 96.72 | 97.05 |
| Mean | 0.90 ± 0.05 | 96.45 ± 0.23 | 96.82 ± 0.23 |
Table 5: Optimal alpha values across folds. Optimal fusion weight alpha determined via grid search (step size 0.05) on validation sets for each cross-validation fold, with mean and SD statistics.
Discusión
Los resultados sugieren que los patrones de activación del código pueden proporcionar información complementaria a los errores de reconstrucción espacial. La mejora del modelo NoCAR al modelo completo FourScales respalda la contribución potencial de la puntuación CAR, mientras que los pesos de fusión seleccionados indican que la rareza semántica podría desempeñar un papel importante en la puntuación final de anomalía. Una posible explicación es que las anomalías relacionadas con el CCR pueden abarcar diferentes escalas morfológicas, desde atipia nuclear hasta la alteración arquitectónica de las glándulas, y por lo tanto podrían beneficiarse de representaciones de características a múltiples niveles de granularidad14. El modelo actual captura diferentes extensiones espaciales mediante fragmentos de 32 × 32 y 64 × 64 a la misma magnificación; la integración futura con marcos de procesamiento de imágenes de toda diapositiva (WSI) podría apoyar aún más un análisis jerárquico a nivel de diapositiva39.
Clínicamente, es preferible considerar a RareCode como una herramienta de triaje previo en lugar de un sistema diagnóstico autónomo. Su alta recuperación (98,40 %) sugiere un riesgo relativamente bajo de pasar por alto casos anómalos40, mientras que su especificidad moderada (58,24 %) probablemente refleje la dificultad de distinguir cambios inflamatorios de tejido normal en un entorno no supervisado. La especificidad más alta específica para cáncer (94,21 %) sugiere además que las falsas alarmas podrían ser menos frecuentes en el subtipo clínicamente más crítico. Estas características de desempeño respaldan el uso potencial de RareCode como herramienta de preselección para priorizar casos que requieren evaluación especializada y agilizar la revisión patológica5,6.
Los hallazgos por clase pueden respaldar aún más el posible papel de RareCode en la triage. El desempeño relativamente superior en cáncer en comparación con la inflamación podría reflejar las alteraciones arquitectónicas y citológicas más pronunciadas que suelen observarse en tejido maligno, incluyendo pérdida de polaridad, pleomorfismo nuclear y desmoplasia estromal14,39. En contraste, los cambios inflamatorios pueden solaparse más estrechamente con la variación tisular benigna, lo que podría explicar en parte la menor especificidad específica para inflamación. Por lo tanto, la especificidad general debe interpretarse con precaución, considerando la detección de cáncer como un caso de uso clínicamente importante más que como un diagnóstico autónomo definitivo.
Los posibles modos de fallo de RareCode deben interpretarse en el contexto de su objetivo de aprendizaje de una sola clase. Dado que el modelo aprende patrones de tejido normal en lugar de categorías específicas de enfermedad, cambios no neoplásicos como epitelio regenerativo, fibrosis, necrosis o inflamación marcada también pueden recibir puntuaciones de anomalía altas. Por el contrario, displasia sutil o carcinoma bien diferenciado con arquitectura glandular casi normal puede producir respuestas más débiles. Factores técnicos, incluyendo variaciones en la tinción, pliegues en el tejido, artefactos de corte, desenfoque y compresión de la imagen, también pueden influir en los errores de reconstrucción y en las frecuencias de activación del código. Estas consideraciones sugieren que la traducción clínica requeriría control de calidad de la imagen, armonización de tinciones, calibración ajustada al escáner y validación multicéntrica39,41.
Alcance de las comparaciones de referencia. La evaluación experimental de este estudio se centra en métodos que operan bajo supuestos comparables de datos y capacidad computacional, específicamente aquellos que pueden entrenarse de extremo a extremo utilizando únicamente muestras normales no etiquetadas a pequeña escala, sin requerir corpus externos de preentrenamiento. Este alcance incluye los principales paradigmas en detección no supervisada de anomalías: basados en reconstrucción (CAE, VAE, SAE, MAE, PatchSAE), aumentados con memoria (MemAE), por destilación de conocimiento (STFPM), por aumentación sintética (DRAEM) y por coincidencia de características previamente entrenadas (PatchCore). Reconocemos que dos categorías de métodos no se incluyeron como líneas base directas. Primero, los modelos fundamentales específicos de patología (UNI, CONCH, CTransPath) aprovechan el preentrenamiento con millones de imágenes de patología curadas, y sus ventajas de desempeño se deben principalmente a la escala y diversidad de los datos de preentrenamiento, más que al mecanismo de detección de anomalías; por lo tanto, una comparación directa confundiría las contribuciones de los datos de preentrenamiento con las del método de detección. El desempeño subóptimo de STFPM y PatchCore, ambos que utilizan estructuras previamente entrenadas en ImageNet, ilustra empíricamente el impacto de la brecha de dominio cuando se aplican características previamente entrenadas de propósito general a escenarios histopatológicos, lo que sugiere que el preentrenamiento específico de patología podría cerrar esta brecha. Segundo, los métodos de detección de anomalías basados en difusión, aunque prometedores, imponen una sobrecarga computacional considerablemente mayor durante la inferencia (normalmente requieren cientos de pasos iterativos de eliminación de ruido), lo que limita su aplicabilidad en flujos de trabajo clínicos de cribado de alto rendimiento, donde la eficiencia del procesamiento es un requisito práctico. En trabajos futuros se investigará si la incorporación de codificadores previamente entrenados específicos de patología en la arquitectura RareCode—reemplazando el codificador actual entrenado de extremo a extremo—puede mejorar aún más el desempeño de detección, manteniendo al mismo tiempo las ventajas de interpretabilidad del mecanismo CAR.
Se deben reconocer varias limitaciones. Primero, la validación se limitó a un conjunto de datos de CCR de un solo centro, y se necesita una evaluación multicéntrica adicional entre instituciones, escáneres y protocolos de tinción. Segundo, aunque la especificidad específica para el cáncer fue alentadora, la especificidad general siguió afectada por el límite entre inflamación y tejido normal. Tercero, no se dispuso de anotaciones expertas a nivel de píxel; por lo tanto, no fue posible calcular los índices de Dice y de IoU, y la evaluación de localización se limitó a la visualización cualitativa y al análisis a nivel de imagen de las respuestas espaciales agregadas. Cuarto, aún se necesitan estudios formales con lectores para determinar si los mapas de calor jerárquicos pueden mejorar la precisión diagnóstica o la eficiencia de revisión. Finalmente, RareCode actualmente realiza análisis a nivel de fragmento e imagen, y la implementación a nivel de WSI requeriría la integración con marcos adicionales de procesamiento39,41,42.
Trabajos futuros deben evaluar RareCode en bancos de pruebas públicos y cohortes multicéntricas tras su integración con marcos de procesamiento de WSI. Estudios prospectivos con lectores podrían ayudar a evaluar su impacto potencial en la precisión diagnóstica, el tiempo de revisión y el acuerdo entre observadores42. Otras líneas de investigación incluyen mejorar la eficiencia computacional, extender el modelo hacia la subtipificación multiclase de anomalías e integrar el aprendizaje basado en múltiples instancias para el diagnóstico a nivel de portaobjetos28,39.
Divulgaciones
Los autores declaran que no tienen intereses financieros competitivos conocidos ni relaciones personales que pudieran haber parecido influir en el trabajo descrito en este artículo. Durante la elaboración de este trabajo, utilizamos ChatGPT para mejorar el lenguaje y la legibilidad del manuscrito y para ayudar en la generación de código de visualización de datos. Tras utilizar esta herramienta, revisamos y editamos el contenido según fue necesario y asumimos toda la responsabilidad por el artículo publicado.
Agradecimientos
Esta investigación fue apoyada por la Estación de Trabajo del Académico Yong Dai (Tecnología Diagnóstica para Enfermedades Autoinmunes) (Ciencia y Tecnología Wan Ke [2023] N.º 317), el Proyecto del Fondo de Innovación para Posgrados del Centro Nacional de Investigación de Hefei para el Centro Conjunto de Investigación en Medicina Ocupacional y Salud del Fondo Abierto para Ciencia y Tecnología de la Salud (N.º OMH-2023-04) y el Proyecto de Investigación Clínica y Traduccional de la Provincia de Anhui (N.º 202427610020132).
Materiales
| Nombre | Empresa | Número de catálogo | Comentarios |
|---|---|---|---|
| GPU GeForce RTX 4060 Ti | NVIDIA | N/A | 16 GB de VRAM; se utilizó una sola GPU para todos los experimentos (RRID:SCR_022858) |
| Matplotlib | Equipo de desarrollo de Matplotlib | Versión 3.8.4 | Visualización de datos y generación de figuras (RRID:SCR_008624) |
| NumPy | Desarrolladores de NumPy | Versión 1.26.4 | Computación numérica y operaciones con arreglos (RRID:SCR_008633) |
| pandas | Equipo de desarrollo de pandas | Versión 2.2.3 | Organización de datos y procesamiento de resultados tabulares (RRID:SCR_018214) |
| Pillow | Colaboradores de Python Imaging Library / Pillow | Versión 10.3.0 | Carga y preprocesamiento de imágenes |
| Python | Fundación Python Software Foundation | Versión 3.12.3 | Lenguaje de programación utilizado para el desarrollo del modelo y el análisis de datos (RRID:SCR_008394) |
| PyTorch | Meta Platforms, Inc. | Versión 2.7.0+cu118 | Entorno de aprendizaje profundo utilizado para la implementación y entrenamiento del modelo (RRID:SCR_018536) |
| scikit-learn | Desarrolladores de scikit-learn | Versión 1.4.2 | Validación cruzada, división entre entrenamiento y prueba, y métricas de evaluación (RRID:SCR_002577) |
| SciPy | Desarrolladores de SciPy | Versión 1.13.1 | Análisis estadístico (RRID:SCR_008058) |
| torchvision | Proyecto PyTorch | Versión 0.22.0+cu118 | Preprocesamiento de imágenes y transformación de datos |
| tqdm | Desarrolladores de tqdm | Versión 4.66.4 | Supervisión del progreso durante el entrenamiento y la evaluación |
Referencias
- Fan J, Sun Q, Di Y, et al. DIPathMamba: A domain-incremental weakly supervised state space model for pathology image segmentation. Med Image Anal. 2025;103:103563.
- Nan T, Zheng S, Qiao S, et al. Deep learning quantifies pathologists' visual patterns for whole slide image diagnosis. Nat Commun. 2025;16(1):5493.
- Lagogiannis I, Meissen F, Kaissis G, et al. Unsupervised pathology detection: a deep dive into the state of the art. IEEE Trans Med Imaging. 2024;43(1):241-252.
- Tian Y, Liu F, Pang G, et al. Self-supervised pseudo multiclass pre-training for unsupervised anomaly detection and segmentation in medical images. Med Image Anal. 2023;90:102930.
- Metter DM, Colgan TJ, Leung ST, et al. Trends in the US and Canadian pathologist workforces from 2007 to 2017. JAMA Netw Open. 2019;2(5):e194337.
- Shaukat A, Levin TR. Current and future colorectal cancer screening strategies. Nat Rev Gastroenterol Hepatol. 2022;19(8):521-531.
- Ancker JS, Edwards A, Nosal S, et al. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Med Inform Decis Mak. 2017;17(1):36.
- Litjens G, Kooi T, Ehteshami Bejnordi B, et al. A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60-88.
- Stepec D, Skocaj D. Unsupervised detection of cancerous regions in histology imagery using image-to-image translation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. 2021:3785-3792.
- Li Y, Lao Q, Kang Q, et al. Self-supervised anomaly detection, staging and segmentation for retinal images. Med Image Anal. 2023;87:102805.
- Hinton GE, Salakhutdinov RR. Reducing the dimensionality of data with neural networks. Science. 2006;313(5786):504-507.
- Kingma DP, Welling M. Auto-encoding variational Bayes. In: Proceedings of the 2nd International Conference on Learning Representations (ICLR). 2014.
- Goodfellow IJ, Pouget-Abadie J, Mirza M, et al. Generative adversarial nets. In: Advances in Neural Information Processing Systems. 2014;27:2672-2680.
- Lu S, Zhang W, Zhao H, et al. Anomaly detection for medical images using heterogeneous auto-encoder. IEEE Trans Image Process. 2024;33:2770-2782.
- Yang Z, Zhang T, Soltani Bozchalooi I, et al. Memory-augmented generative adversarial networks for anomaly detection. IEEE Trans Neural Netw Learn Syst. 2022;33(6):2324-2334.
- Huyan N, Quan D, Zhang X, et al. Unsupervised outlier detection using memory and contrastive learning. IEEE Trans Image Process. 2022;31:6440-6454.
- Jézéquel L, Beaudet J, Histace A, et al. Unified anomaly detection via multi-scale contrasted memory. IEEE Trans Image Process. 2026;35:2802-2815.
- Roth K, Pemula L, Zepeda J, et al. Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022:14298-14308.
- Defard T, Setkov A, Loesch A, et al. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In: International Conference on Pattern Recognition. Cham: Springer; 2021:475-489.
- Zingman I, Stierstorfer B, Lempp C, et al. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development. Med Image Anal. 2024;92:103067.
- Jiang Y, Cao Y, Shen W. Prototypical learning guided context-aware segmentation network for few-shot anomaly detection. IEEE Trans Neural Netw Learn Syst. 2025;36(7):12016-12026.
- Wang X, Zhao J, Marostica E, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature. 2024;634(8035):970-978.
- Frotscher A, Kapoor J, Wolfers T, et al. Unsupervised anomaly detection in medical imaging using aggregated normative diffusion. Med Image Anal. 2026;109:103895.
- Zeng GQ, Yang YW, Lu KD, et al. Evolutionary adversarial autoencoder for unsupervised anomaly detection of industrial Internet of Things. IEEE Trans Reliab. 2025;74(3):3454-3468.
- Lu KD, Zhang BX, Xu Y, et al. MoARNN-AM: multi-objective automated recurrent neural network with attention mechanism for cyber-attack detection of UAV. IEEE Trans Consum Electron. 2026;72(1):1738-1749.
- van den Oord A, Vinyals O, Kavukcuoglu K. Neural discrete representation learning. In: Advances in Neural Information Processing Systems. 2017;30:6306-6315.
- Ghafourian A, Shui H, Upadhyay D, et al. Targeted collapse regularized autoencoder for anomaly detection: black hole at the center. IEEE Trans Neural Netw Learn Syst. 2025;36(6):10348-10358.
- Gao Z, Mao A, Dong Y, et al. SMMILe enables accurate spatial quantification in digital pathology using multiple-instance learning. Nat Cancer. 2025;6(12):2025-2041.
- Wang X, Liu H, Zhang Y, et al. Joint modeling histology and molecular markers for cancer classification. Med Image Anal. 2025;102:103505.
- Jin H, Shen J, Cui L, et al. Dynamic graph based weakly supervised deep hashing for whole slide image classification and retrieval. Med Image Anal. 2025;101:103468.
- Schmitz R, Madesta F, Nielsen M, et al. Multi-scale fully convolutional neural networks for histopathology image segmentation: from nuclear aberrations to the global tissue architecture. Med Image Anal. 2021;70:101996.
- Masci J, Meier U, Ciresan D, et al. Stacked convolutional auto-encoders for hierarchical feature extraction. In: International Conference on Artificial Neural Networks. Berlin: Springer; 2011:52-59.
- Ng A. Sparse autoencoder. CS294A Lecture Notes. 2011;72:1-19.
- He K, Chen X, Xie S, et al. Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022:16000-16009.
- Lim H, Choi J, Choo J, et al. Sparse autoencoders reveal selective remapping of visual concepts during adaptation. In: Proceedings of the 13th International Conference on Learning Representations (ICLR). 2025.
- Gong D, Liu L, Le V, et al. Memorizing normality to detect anomaly: memory-augmented deep autoencoder for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019:1705-1714.
- Wang G, Han S, Ding E, et al. Student-teacher feature pyramid matching for anomaly detection. In: Proceedings of the 32nd British Machine Vision Conference (BMVC). 2021.
- Zavrtanik V, Kristan M, Skocaj D. DRAEM: a discriminatively trained reconstruction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. 2021:8330-8339.
- Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and artificial intelligence. Lancet Oncol. 2019;20(5):e253-e261.
- Raab SS, Grzybicki DM, Janosky JE, et al. Clinical impact and frequency of anatomic pathology errors in cancer diagnoses. Cancer. 2005;104(10):2205-2213.
- Schömig-Markiefka B, Pryalukhin A, Hulla W, et al. Quality control stress test for deep learning-based diagnostic model in digital pathology. Mod Pathol. 2021;34(12):2098-2108.
- Steiner DF, MacDonald R, Liu Y, et al. Impact of deep learning assistance on the histopathologic review of lymph nodes for metastatic breast cancer. Am J Surg Pathol. 2018;42(12):1636-1646.