Adapting DeepLabV3+ for biopsy cervical cancer lesion segmentation

Area of Science:

• Digital pathology
• Machine learning
• Oncology

Background:

• Cervical cancer is a major cause of mortality in resource-limited settings.
• Limited access to advanced digital pathology equipment hinders diagnosis.
• Automated histopathological image segmentation offers a potential solution.

Purpose of the Study:

• To develop and validate a smartphone-assisted microscopy system for cervical cancer lesion segmentation.
• To assess the performance of the DeepLabV3+ architecture in resource-constrained environments.

Main Methods:

• Utilized smartphone-assisted microscopy with a custom adapter and Ocular app for image acquisition.
• Developed a DeepLabV3+ model with a ResNet34 encoder for segmenting 21 histopathological feature classes.
• Trained the model on 5,966 H&E-stained images from the Uganda Cancer Institute using combined BCE and Dice loss.

Main Results:

• Achieved a mean Intersection over Union (IoU) of 75.8% and a Dice coefficient of 93.1% on a validation set.
• Demonstrated consistent per-class IoU between 74.13% and 75.41% across all feature classes.
• DeepLabV3+ significantly outperformed a U-Net baseline (mIoU: 56.84%, Dice: 68.53%).

Conclusions:

• The smartphone-based system is technically feasible for reliable digital pathology in resource-limited settings.
• DeepLabV3+ architecture effectively delineates complex histopathological patterns.
• Further multi-institutional evaluation is required before clinical deployment.