HistoPath AI

Cancer survival depends heavily on catching it early, yet histopathologists - the specialists who diagnose it from tissue samples - are in increasingly short supply. HistoPath AI is a decision-support tool for that gap: it reads a histopathology image, classifies the tissue as benign or malignant, and is intended to take confident cases off a pathologist’s plate while flagging uncertain ones for human review.

I framed the task as binary classification and built up from honest baselines: a random guess, a simple MLP at around 80%, and an estimate of human-level performance at about 83%. A simple CNN reached 82%; data augmentation lifted it to 93%; VGG16 transfer learning reached 97%; and EfficientNetB0 with augmentation reached about 98%.

The final model misclassified one held-out image - a benign sample flagged as malignant, the safer direction to err in screening. Grad-CAM checks showed that its attention generally aligned with the regions a pathologist would examine.

A model is only half of a medical tool. I designed a clinical Figma prototype with a heatmap beside each prediction, a confidence score, clear assistive-use disclaimers, and medical-ID-gated access. A think-aloud study with four participants validated the core flow and surfaced concrete navigation and labeling improvements.

I also compared interface versions with and without Grad-CAM and a persistent menu across six usability measures. None of the differences were statistically significant, so I reported the null result plainly: one version scored marginally higher, but not meaningfully so.