Estimating Tumor Size on 2D Images of Non-Uniform Body Cavity with Object Detection and Depth Analysis

Tumor size estimation during cancer diagnosis is a determining factor of disease progression. Traditionally, this estimation relies on 3D volumetric imaging techniques, which are essential because of the complex non-uniform structure of body cavities. These imaging techniques could comprehensively map the location of the tumors and provide estimation to tumor sizes. However, such imaging is expensive, leading to a high cost for cancer diagnosis. In contrast, 2D video streams obtained from exploratory surgeries are more accessible and cost effective. However, these 2D streams lack spatial and depth information for reliable tumor size estimation because the non-uniformity of cavities can distort the appearance of tumors. This makes 2D data less explored by studies on tumor size estimation and presents an opportunity for this work to overcome the limitations of 2D data. The first step in this work is to detect the tumors to identify the regions of interest, which is performed using You Only Look Once (YOLO), a deep learning-based object detection method. Next, to address the lack of depth information, DepthAnythingV2 is applied for depth analysis, enabling the accurate resizing of tumors based on their proximity to the camera. The algorithm was chosen due to its generalizability without needing specific training. The limited spatial context is then overcome using video-to-composite image stitching, allowing a more thorough analysis of a single inference. This study used laparoscopic data of peritoneal carcinomatosis through collaboration in clinical settings. The data was manually annotated by surgeons to reflect real-world diagnostic conditions. Despite the limited dataset size, data augmentation was applied by modifying image dimensions to enhance variability. Our findings show that YOLOv11S with batch size 4 and image size 640 achieved the best performance with 0.959 area under precision-recall curve (AUPRC), which could identify tumors with high confidence scores. The addition of depth data improved the prediction of the largest tumor within images with an accuracy of 84.29%, compared to 35.00% without depth.