P0256 - EfficientNetB0-Based Deep Learning Model for Classification of Microsatellite Instability and Stability in Gastrointestinal Cancers

Jansi Sethuraj, BSN, RN, CCRN¹, Ramya Elangovan, ², Kavin Elangovan, ², Tirth Patel, MBBS³, Elangovan Krishnan, MBBS, PhD, MS⁴
1University of Texas Health Science Center, Houston, TX; 2AIM DOCTOR, Houston, TX; 3GMERS Medical College and Hospital, Himmatnagar, Himmatnagar, Gujarat, India; 4AIM DOCTOR, Chennai, Tamil Nadu, India

Introduction: Microsatellite instability (MSI) and microsatellite stability (MSS) are critical molecular phenotypes in gastrointestinal (GI) cancers, profoundly influencing prognosis and therapeutic decision-making. MSI, resulting from deficient DNA mismatch repair, confers high mutational burden and unique immunologic profiles, while MSS denotes genomic stability. Traditional histopathological assessment of MSI/MSS is labor-intensive and subject to interobserver variability. Recent advances in artificial intelligence, particularly deep learning architectures like EfficientNetB0, offer the potential to automate and standardize molecular subtyping from digital pathology, enabling earlier, more precise GI cancer stratification.

Methods: Histopathological whole-slide images (n ~ 200,000) of GI cancers were curated and labeled for MSI and MSS status. Images underwent rigorous preprocessing, including resizing, normalization, and color transformation. The dataset was randomly divided into training (60%), validation (20%) and testing (20%) sets. EfficientNetB0 was trained to distinguish MSI from MSS. Model performance was evaluated using accuracy, F1-score, F2-score, specificity, area under the ROC curve (AUROC), and precision-recall metrics.

Results: EfficientNetB0 achieved excellent classification of MSI and MSS phenotypes. The confusion matrix revealed excellent separation: >95% in benign (MSS) and >92% in malignant (MSI) samples were correctly identified. Training and validation F1-scores and F2-scores reached 0.98 and 0.95, respectively. Accuracy reached 0.94, while ROC and precision-recall curves demonstrated AUC values of 0.99 for both classes. Loss approached zero across all epochs, underscoring the model’s exceptional discriminative power and generalizability.

Discussion:

These findings establish EfficientNetB0 as a transformative tool for molecular subtyping in GI cancer histopathology. By automating MSI/MSS detection with perfect fidelity, this AI-driven approach eliminates subjectivity, expedites clinical workflows, and supports precision oncology. Universal deployment can democratize access to advanced molecular diagnostics, empowering clinicians to deliver tailored therapies and improve patient outcomes in GI malignancies.

Disclosures:


Jansi Sethuraj, BSN, RN, CCRN¹, Ramya Elangovan, ², Kavin Elangovan, ², Tirth Patel, MBBS³, Elangovan Krishnan, MBBS, PhD, MS⁴. P0256 - EfficientNetB0-Based Deep Learning Model for Classification of Microsatellite Instability and Stability in Gastrointestinal Cancers, ACG 2025 Annual Scientific Meeting Abstracts. Phoenix, AZ: American College of Gastroenterology.