melanoma_Yolo

Melanoma Lesion Detection — YOLOv8n

Single-class object detection model that localizes skin lesions in dermoscopic images. This is the detection stage of a 3-model melanoma analysis pipeline (detection → segmentation → classification).

• Architecture: YOLOv8n (nano)
• Task: single-class detection (0 = lesion)
• Training data: ISIC 2018 Task 1, segmentation masks converted to bounding boxes
• Result: mAP50 = 0.977
• Code: github.com/AdamSobanski/melanoma_yolo

Role in the pipeline

This model is the fast first pass in a larger pipeline: it crops the lesion region out of the full dermoscopic image before handing it off to a U-Net segmentation model and an EfficientNet-B4 classifier. Full pipeline: melanoma_pipline_yolo_seg_class_v2.

How the training data was built

ISIC 2018 Task 1 provides pixel-level segmentation masks, not bounding boxes. Each mask was converted to a single YOLO-format bounding box by taking the bounding rectangle of the lesion region (min/max row and column of the thresholded mask), then normalized to x_center, y_center, width, height.

Training details

• Base weights: yolov8n.pt (Ultralytics pretrained)
• Epochs: 10
• Image size: 640
• Batch size: 16
• Hardware: Kaggle, single GPU (T4)

from ultralytics import YOLO

model = YOLO('yolov8n.pt')
model.train(
    data='yolo_dataset/dataset.yaml',
    epochs=10,
    imgsz=640,
    batch=16,
    device=0,
)