Application of convolutional neural networks for histopathological diagnosis of feline low-grade T-cell lymphoma and lymphoplasmacytic enteritis in intestinal biopsies

Le résumé fourni par la source

Introduction: Low-grade intestinal T-cell lymphoma (LGITL) and lymphoplasmacytic enteritis (LPE) are common causes of feline chronic enteropathy. Although histopathology is the diagnostic gold standard, distinguishing between LGITL and LPE is challenging due to overlapping morphological features and high interobserver variability. Convolutional Neural Networks (CNNs) may improve diagnostic consistency and accuracy. Methods: This proof-of-concept study retrospectively analysed 161 formalin-fixed, paraffin-embedded endoscopic intestinal biopsies from cats diagnosed with LGITL or LPE. Samples were blindly re-evaluated according to standard guidelines using haematoxylin-eosin (H&E) and immunohistochemistry (IHC) by two board-certified veterinary pathologists; discordant cases were excluded. An InceptionV3 CNN, trained via transfer learning, was applied to 8,026 manually selected image tiles (1,024 × 1,024 pixels, RGB) obtained from selected digitized H&E-stained sections. Training included tile balancing and image augmentation, and the network used a 5-fold cross-validation strategy. The same cases were independently reviewed by three board-certified pathologists for comparison. Results: The final dataset comprised 142 cases, including 104 LPE and 38 LGITL. A test set of 23 cases was classified by the CNN, which generated a final diagnosis based on majority tile vote. The CNN achieved an average tile-level accuracy of 85.3% in cross-validation, focusing attention on lymphocyte-rich areas, as indicated by Grad-CAM analysis. At the case level, using a tile majority vote system, the CNN correctly diagnosed ~95% of 23 test cases in 1 min and 20 s, compared to pathologists' average accuracy of ~85% in 9 min. Notably, the only case misclassified by the CNN was also misdiagnosed by two of the three pathologists. Discussion: These findings suggest that CNN-based analysis has strong potential as a diagnostic support tool for differentiating LGITL from LPE. Further model optimization and dataset expansion may enhance performance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.