1 of 10

MACHINE LEARNING | PROJECT

PolypVision AI

Automated Polyp Detection

Computer-aided polyp detection for clinicians

Vinod Kumar Prajapat | BT23ECE122 ·  Apoorv Deshmukh | BT23ECE013

Dept. of Electronics & Communication | ML Project

9,035

Dataset Images

2.3M

YOLOv11n Params

2 of 10

INTRODUCTION

Introduction & Objective

The Clinical Problem

20–25%

Polyp miss rate in routine colonoscopies.

  • High endoscopist workload and fatigue
  • Subtle early-stage polyps are easy to miss

Our Objective

Build a clinical AI system for colonoscopy images.

  • From-scratch YOLOv11 implementation
  • Designed for polyp morphology and real-time screening support

"A secondary AI eye in the room — so no polyp goes unnoticed."

3 of 10

DATA

Data Collection & Preprocessing

Data Collection

Images from multiple sources provide diversity in morphology, lighting, and equipment.

[Roboflow] polyp-kntak

[Roboflow] polyp-detection-xdae2

9,035

Total Images — 100 %

6,502

Training Set ~ 72 %

902

Validation Set ~ 10%

1,631

Test Set ~ 18%

Data Pipeline

Input

RGB images, 224×224 to 1920×1080

Letterbox

Resize to 640×640, pad gray 114

Augmentation

HSV jitter; flip 50%

Images are letterboxed to 640×640 with gray padding (114), preserving aspect ratio. Training uses HSV jitter every batch and 50% horizontal flip.

~884K unique visual inputs — 136 epochs × 6,502 training images.

4 of 10

ARCHITECTURE

YOLOv11n Architecture — Built From Scratch

~2.3M parameters — a lightweight detection backbone for medical imagery.

5 of 10

6 of 10

TRAINING

Training Pipeline

Optimization Strategy

  • AdamW optimizer with decoupled weight decay

​

  • Cosine annealing LR schedule

​

  • 3-epoch linear warmup

​

  • 300 epochs; early stopping at 50 epochs

Loss Functions

Task

Loss Function

Classification

BCE + TAL Assigner

Bounding Box

CIoU + DFL

Hardware: NVIDIA RTX GPU

7 of 10

RESULTS

Evaluation Results — Test Set Performance

Evaluated on 1,631 independent test images held out during development.

0.9343

mAP@0.5

Mean Average Precision at IoU 0.5

0.9749

Precision

True positive rate of predicted polyps

0.9299

Recall

Detected actual polyps

0.9519

F1 Score

Precision-recall balance

CPU Inference

~10.48 FPS

8 of 10

9 of 10

FUTURE WORK

Future Directions

Model Pruning

Cut parameters for mobile edge deployment and point-of-care use without GPU infrastructure.

Grad-CAM Explainability

Show saliency heatmaps that highlight the image regions driving each prediction.

Multi-Class Extension

Classify polyp subtypes—adenomatous vs. hyperplastic—for better risk stratification.

Multi-View Frame Fusion

Fuse sequential frames to reduce flicker and improve temporal consistency in live colonoscopy.

10 of 10

CONCLUSION

Summary & Impact

Built from Scratch

Created a full YOLOv11n pipeline independently, with complete architectural control.

Clinical Results

Reached 0.9343 mAP@0.5 and 0.9749 Precision on 1,631 test images from real colonoscopy data.

Clinical Value

Supports early colorectal cancer detection by helping ensure polyps are not missed.

"A high-accuracy model for early detection."

Vinod (BT23ECE122) · Apoorv (BT23ECE013) | Dept. of Electronics & Communication | ML Project