Smartphone-Based Oral Lesion Image Segmentation Using Deep Learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Early detection of oral diseases is critical for improving patient outcomes.
• Accurate segmentation of oral lesions aids clinicians and enhances deep learning (DL) diagnostic models.
• Smartphone-based imaging offers a potential avenue for accessible oral disease screening.

Purpose of the Study:

• To develop a deep learning (DL) solution for segmenting oral lesions from smartphone-captured images.
• To design and evaluate a novel UNet-based model, OralSegNet, for enhanced segmentation accuracy.
• To provide a cost-effective and non-invasive tool for early oral disease diagnosis.

Main Methods:

• A novel UNet-based model, OralSegNet, was designed using EfficientNetV2L as the encoder, incorporating Atrous Spatial Pyramid Pooling (ASPP) and residual blocks.
• A dataset of 538 oral lesion images was utilized, pre-processed, resized to 256x256 pixels, and augmented for robustness.
• The model's performance was evaluated using Dice coefficients and Intersection over Union (IoU) scores on validation and test sets.

Main Results:

• OralSegNet achieved high performance with Dice coefficients of 0.9530 (validation) and 0.8518 (test), and IoU scores of 0.9104 (validation) and 0.7550 (test).
• The model outperformed traditional and state-of-the-art segmentation models.
• Despite its parameter count, OralSegNet demonstrated computational efficiency with the lowest FLOPS (34.30 GFLOPs).

Conclusions:

• OralSegNet provides an accurate and efficient deep learning-based solution for oral lesion segmentation from smartphone images.
• The model's performance and cost-effectiveness make it a valuable tool for clinicians, facilitating early diagnosis.
• This technology has the potential to improve accessibility to oral disease diagnosis, particularly in rural or resource-limited settings.