Authors:

Yagnesh Lokesh

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Early detection and effective treatment of Uveal Melanoma (UM), one of the most common intraocular malignancies in adults, relies on accurate tumor measurement. Ultrasonography is the gold-standard for measuring choroidal tumors, but manual segmentation is highly operator-dependent, subject to inter-.and intra-observer variability, and time-consuming. This study explores the potential of deep learning as a tool for automated tumor segmentation and measurement. A DenseNet-121 U-net model was trained to segment three classes on 114 B-scan ultrasound images. Ground-truth masks were annotated with tumor (green), inner retina (red), and inner scleral (blue) boundaries. The model was trained using a combined cross-entropy and Dice loss, and performance was evaluated using Dice score, Average Symmetric Surface Distance (ASSD), and 95th percentile Hausdorff Distance (HD95). The model achieved an average Dice score of 0.91 at the tumor margins, an ASSD of 3.5 pixels, and an HD95 of 17.8 pixels. The inner retinal lines were segmented with an ASSD of 1.7 pixels and an HD95 of 12.5 pixels, while the inner scleral boundary had an ASSD of 3.5 pixels and an HD95 of 19.9 pixels. This study shows that a deep learning model can accurately segment relevant anatomical layers and tumor boundaries in ultrasound images, as evidenced by the high Dice score and low boundary error in the tumor region. While the inner scleral boundary showed higher errors than the other structures, the overall segmentation performance remained strong. These findings suggest that deep learning based segmentation could support early detection of choroidal tumors, particularly in clinical settings with limited access to ocular oncologists. Further training with larger datasets and external validation will help confirm generalizability and support future clinical implementation.

Source:

University of Illinois Chicago

Topics:

No topics listed

Co-authors:

Yagnesh Lokesh