Decoding Future Risk: Deep Learning Analysis of Tubular Adenoma Whole-Slide Images

Colorectal cancer (CRC) remains a significant cause of cancer-related mortality, despite the widespread implementation of prophylactic initiatives aimed at detecting and removing precancerous polyps. Although screening effectively reduces incidence, a notable portion of patients initially diagnosed with low-grade adenomatous polyps will still develop CRC later in life, even without the presence of known high-risk syndromes. Identifying which low-risk patients are at higher risk of progression is a critical unmet need for tailored surveillance and preventative therapeutic strategies. Traditional histological assessment of adenomas, while fundamental, may not fully capture subtle architectural or cytological features indicative of malignant potential. Advancements in digital pathology and machine learning provide an opportunity to analyze whole-slide images (WSIs) comprehensively and objectively. This study investigates whether machine learning algorithms, specifically convolutional neural networks (CNNs), can detect subtle histological features in WSIs of low-grade tubular adenomas that are predictive of a patient’s long-term risk of developing colorectal cancer.

💡 Research Summary

This paper investigates whether deep learning can uncover subtle histopathologic cues in low‑grade tubular adenomas (TAs) that predict a patient’s future risk of developing colorectal cancer (CRC). The authors performed a retrospective cohort study of patients who underwent screening colonoscopy between 2013 and 2022 and had biopsy‑confirmed low‑grade tubular adenomas. Patients with known high‑risk conditions (IBD, Lynch syndrome, FAP, prior CRC, or high‑grade dysplasia) were excluded. The remaining cohort was divided into “progressors” (those who later developed CRC, n = 32) and “non‑progressors” (no CRC, n = 22). Clinical data showed that progressors were older (median 79 vs 69 years), had longer intervals between screenings, and underwent fewer colonoscopies, reflecting real‑world surveillance disparities.

All H&E‑stained slides were digitized at 40× magnification (Leica Aperio AT2) and stored as SVS whole‑slide images (WSIs). Expert pathologists annotated adenomatous epithelium and exclusion zones (normal mucosa, lymphoid aggregates, artifacts) using QuPath. After quality control, WSIs were tiled into non‑overlapping 1024‑pixel patches, resized to 224 × 224 pixels, color‑normalized, sharpened, and artifact‑filtered, yielding 335,763 high‑quality tiles (143,080 from progressors, 192,683 from non‑progressors). A pre‑trained EfficientNetV2‑Small (EfficientNetV2S) CNN was fine‑tuned on these tiles. Training proceeded in two phases: first, only the final classification layer was trained for 2 epochs (learning rate 0.01); second, the entire network was unfrozen and trained for 25 epochs with a cosine‑decayed learning rate starting at 7 × 10⁻⁵. Data augmentation (Albumentations) introduced random color shifts, saturation, brightness/contrast changes, and sharpening to mimic inter‑lab variability. Dropout (0.2) and early stopping based on validation loss mitigated overfitting; the Adam optimizer was used throughout.

Performance was evaluated at both tile and slide levels. On a held‑out test set of 40,514 tiles, the model achieved an accuracy of 0.9788, precision 0.9762, recall 0.9815, F1‑score 0.9789, and an area under the ROC curve essentially equal to 1.0. The confusion matrix showed 19,773 true negatives, 19,883 true positives, 484 false positives, and 374 false negatives. For whole‑slide inference, 10 progressor and 10 non‑progressor WSIs (total = 20) were completely excluded from training. Tile‑level probabilities were averaged per slide, and a 0.5 threshold yielded correct classification of all 20 slides, demonstrating strong generalization.

Interpretability was explored with Gradient‑Weighted Class Activation Mapping (Grad‑CAM). In progressor tiles, the model highlighted regions with increased architectural complexity (branching, cribriform patterns), nuclear crowding, elongation, and pseudostratification. Non‑progressor tiles emphasized preserved glandular architecture and uniform nuclear spacing. These visual cues suggest that the CNN captures morphologic features that are difficult to quantify manually but correlate with malignant potential.

Statistical analysis of the clinical cohort underscores that older age and less frequent surveillance are associated with progression, yet the deep‑learning signal remains independent of these factors, offering a novel histology‑based risk marker. Limitations include the modest sample size, single‑institution data, and lack of external validation. The authors propose future work incorporating multi‑center cohorts, integrating molecular/genomic data, and conducting prospective trials to assess whether AI‑driven risk stratification can safely extend surveillance intervals for low‑risk patients while intensifying follow‑up for those flagged as high‑risk.

In conclusion, the study provides compelling evidence that low‑grade tubular adenomas harbor machine‑detectable histologic signatures predictive of future CRC. The EfficientNetV2S model achieved near‑perfect discrimination at both tile and slide levels, and Grad‑CAM analysis offers biologically plausible explanations for its decisions. This work paves the way for AI‑augmented pathology to refine post‑polypectomy surveillance, moving toward personalized, data‑driven colorectal cancer prevention.