Nizomiddin Xalilov

Work / Prostate Cancer Detection & Gleason Grading

2023–2024 · Fergana State Technical University · graduation thesis Research

Prostate Cancer Detection & Gleason Grading

A YOLOv5 model that finds cancer regions in prostate biopsy images and grades them on the Gleason scale — in about 0.1 s per image.

My graduation thesis. A pathologist grades prostate cancer by looking at tissue under a microscope, and telling Grade 3 from Grade 4 is hard even for specialists. I built a detector that marks suspicious regions on a histopathology image and assigns each one a class — Benign, Grade 3, Grade 4 or Grade 5 — with a confidence score, as a second opinion for the doctor.

Accuracy on unseen images
89%

97% on the training set

Grade 3 vs Grade 4
85%+

The hardest pair to tell apart

Time per image
~0.1 s
Problem

A decision made by eye

The Gleason grade decides how a patient is treated. It is set by a pathologist reading glass slides, which is slow and subjective: two specialists can disagree on the same sample, especially between Grade 3 (glands still formed) and Grade 4 (glands fused).

Decisions & trade-offs

How it was built

  • Data. 500+ high-resolution biopsy images, each region labelled with one of four classes: Benign, Grade 3, Grade 4, Grade 5.
  • Preparation. Images resized to 640×640 and expanded with mosaic and other augmentations to 1,776 training images.
  • Model. YOLOv5m, chosen because it both locates a region and classifies it in one pass — a doctor needs to see where, not just a yes/no.
  • Validation. Two test sets: 50 images similar to the training data, and a second set of completely new images.
  • App. A Tkinter desktop program: load an image, get the marked regions with class and confidence, and a short written conclusion.
Outcome

Outcome

On images it had never seen, the model reached 89% accuracy. It separates healthy tissue from Grade 5 almost perfectly, and tells Grade 3 from Grade 4 — the pair that troubles pathologists most — with over 85% accuracy. One image takes about 0.1 seconds.

The thesis was defended at Fergana State Technical University and received a grant from the Silk Road Health Data Science community.

What I'd do differently

What I would do differently

The gap between 97% on training images and 89% on new ones is the most honest number in this project: the model learned part of the training set by heart. Today I would collect slides from more laboratories and scanners before training longer, and I would report the confusion matrix first instead of a single accuracy figure.