Intraoperative 3D reconstruction and deep learning-based annotation of cutaneous squamous cell carcinoma from Mohs micrographic surgery
Shritej Mamidala & Abhi Kurapati
Lay Summary:
In our project, Shritej and I were able to detect cancerous regions in tumor belonging to cSCC tumors using images from the Mohs Micrographic surgery. We then aligned the images on top of each other in 3d to aid with the surgery.
Abstract:
Cutaneous squamous cell carcinoma (cSCC) is the second most common skin cancer and often requires Mohs micrographic surgery (MMS) for tumors in areas. While MMS provides complete histopathologic margin control,the procedure is inherently two-dimensional and relies on the coordinated efforts of multiple specialists, a process that risks missed margins and potential overtreatment. To address this limitation, we developed a fully automated pipeline for intraoperative margin assessment that integrates convolutional neural networks–graph neural networks (CNN–GNN)–based classification, rigid registration, and three-dimensional (3D) reconstruction of whole slide images (WSIs). Our dataset comprised 95 hematoxylin and eosin (H&E) stained frozen section slides from cSCC MMS specimens. A ResNet-101 CNN generated patch embeddings, which were embedded into radius-based graphs and classified using a lightweight graph attention network to distinguish benign from malignant tissue. Rigid registration of serial WSIs was performed with VALIS, enabling alignment of tissue islands across sections, and 3D reconstructions were generated to visualize tumor invasion depth and continuity. The CNN–GNN achieved an area under the curve (AUC) of 0.955, consistently outperforming the CNN baseline across all evaluation metrics. Rigid registration reduced target registration error by more than 95% compared with baseline and non-rigid approaches. The entire workflow executed in approximately 10 minutes, demonstrating feasibility for intraoperative integration. Together, these results establish an end-to-end AI-driven framework for 3D tumor reconstruction in MMS, with potential to enhance intraoperative decision-making and improve surgical precision in the management of cSCC.
Q&A:
Bios: Shritej Mamidala,Abhi Kurapati
Program Track: Advanced Research
GitHub Username:
ShritejMamidala -Shritej Mamidala
Abhi-825 -Abhi Kurapati
What was your favorite seminar? Why?
My favorite seminar was by Zarif Azher, who spoke about multimodal modeling in a way that was both highly informative and relevant. I also found his discussion on research and entrepreneurship especially inspiring, as he emphasized how closely the two are connected and how they can complement one another. -Shritej Mamidala
My favorite seminar was actually the first one with Louis Vaickus as the speaker. I enjoyed this seminar because it first went into a topic that I am extremely interested in: AI and Computer Vision. Dr. Vaickus also went on to talk about pathology and the medical operations behind the code, which gave me an informational peek at the medical aspect. Just an overall informative, educational, and interesting seminar. -Abhi Kurapati
If you were to summarize your summer internship experience in one sentence, what would it be?
My summer internship was a valuable and challenging experience that allowed me to learn extensively, contribute to meaningful research, and grow more confident in my abilities and independence. -Shritej Mamidala
I had an amazing experience at the EDIT AI Summer Internship program because I met new people including my team member, delved into extremely educational, informative topics, and worked on a project that was interesting and related to my field of study. -Abhi Kurapati