The Engineering Journey
From Single Model to Multi-Specialist Platform

Version 1.0: Establishing the Baseline (VGG16)

Our initial proof-of-concept utilized a heavyweight VGG16 architecture focused on a binary classification task (Colon Adenocarcinoma vs. Normal). While this 150MB+ model successfully established our foundational Grad-CAM explainability pipeline and achieved high baseline accuracy, it exhibited a common deep learning phenomenon: network overconfidence.

The massive parameter count acting on a simple binary sigmoid output led to mathematical saturation, resulting in "100% confidence" predictions even on out-of-distribution data. In a clinical setting, nuance is critical. We needed a scalpel, not a sledgehammer.

Version 2.0: The Multi-Organ Experiment (MobileNetV2)

We transitioned to a lightweight MobileNetV2 backbone (8.6 MB) and expanded to a 5-class multi-organ model covering both Colon and Lung histopathology. This achieved ~98% accuracy and eliminated the binary overconfidence issue.

However, a critical problem emerged: although the model detected cancer accurately, it sometimes produced organ-mismatched reports — classifying colon tissue with a lung report and vice versa. This cross-organ confusion made the multi-organ approach unreliable for clinical-grade screening and motivated a fundamental rethinking of the model scope.

Version 3.0: The Colon Specialist (EfficientNetB1)

To solve the organ-confusion problem, we adopted a single-organ specialist strategy. The EfficientNetB1 backbone trained exclusively on NCT-CRC-HE-100K and validated on the independent CRC-VAL-HE-7K holdout set — achieving 99.1% internal and 92.7% external accuracy across 9 colorectal tissue types with post-training temperature calibration (T=0.5576).

Version 4.0: Multi-Specialist Platform (Colon + Brain)

V4 transforms X-Pathology from a single-model tool into a multi-specialist diagnostic platform. A new EfficientNetB0 Brain Tumor MRI Specialist joins the Colon Specialist, classifying brain MRI scans into 4 tumor categories. The backend now uses a dynamic specialist routing architecture, loading and dispatching to the correct model based on user selection — proving that the specialist-per-organ paradigm scales cleanly to new medical domains.

The XPathology Brain Tumor MRI Specialist is a fine-tuned EfficientNetB0 CNN trained to classify T1-weighted axial brain MRI scans into 4 diagnostic categories: glioma, meningioma, pituitary tumor, and no tumor. Trained on the Brain Tumor MRI Dataset (Masoud Nickparvar, Kaggle), it uses the dataset's standard Training/Testing split with a 1,600-image holdout set that was never seen during training.

The model uses a two-phase training strategy: a 10-epoch warm-up with a frozen backbone (LR=1e-3), followed by fine-tuning with the top 20% of layers unfrozen (LR=1e-5). MRI-specific augmentations include horizontal flips (brain symmetry), ±15° rotation, brightness/contrast variation — but no vertical flip or hue/saturation changes, as these are anatomically invalid for MRI.

Built By

AgenticEra Systems

Pioneering the intersection of agentic AI and real-world applications. From intelligent code assistants to medical diagnostics — we build AI systems that augment human expertise with precision and responsibility.