KEC Conference 2025InJET PublishedDOI

Automated NucleiSegmentation

Deep learning U-Net model for automated cell nucleus detection in microscopy images. Upload your image and get instant segmentation masks with confidence heatmaps.

97.5%
Accuracy
0.88
IoU Score
~99ms
Inference
1.94M
Parameters

Or try a sample image

Architecture

U-Net Architecture

Symmetric encoder-decoder with skip connections. 1.94M trainable parameters. Input 256×256 RGB → Output 256×256 binary mask.

↓16256px↓32128px↓6464px12832px↓25616pxbottleneck↑12832px↑6464px↑32128px↑16256pxInput256×256×3Output256×256×1EncoderDecoderSkip connection~1.94M params · 5 encoder + 4 decoder blocks · ReLU + Dropout · He Normal init

Performance

Comparison with State-of-the-Art

ModelAccuracyPrecisionRecallIoU
DeepLabV3+91.2%88.7%89.1%80.5%
DeepResNet94.6%88.6%81.7%85.3%
Proposed U-NetOURS97.5%94.5%95.1%88.2%

Robustness Across Dataset Splits

SplitTrainValTestIoU
Original67010600.88
80-10-1059274740.85
70-20-10518148740.83

Preprocessing Pipeline

  1. 1. Resize to 256×256 (LANCZOS)
  2. 2. Gaussian filter σ≈1.0 (noise reduction)
  3. 3. Normalize pixel values ÷ 255
  4. 4. Forward pass through U-Net
  5. 5. Threshold at 0.5 → binary mask
  6. 6. Count connected components

Explainability

Grad-CAM — attention heatmaps showing which regions influenced predictions. Consistently highlights nuclear boundaries (red) vs background (blue).
LRP — pixel-wise relevance propagation. Mean relevance 0.45, confidence score 0.902. Confirms model focuses on biologically meaningful nuclear features.

Dataset

2018 Kaggle Data Science Bowl nuclei dataset — 740 microscopy images (670 train / 10 val / 60 test). Stratified sampling across imaging modalities and nuclear densities. Mean 71±9 nuclei per image. ~27% of images contain significant nuclear overlap.