Enhancing lesion detection in liver and kidney CT scans via lesion mask selection from two models: A main model and a model focused on small lesions.

Automated segmentation and detection of tumors in CT scans of the liver and kidney have a significant potential in assisting clinicians with cancer diagnosis and treatment planning. However, current approaches, including state-of-the-art deep learning ones, still face many challenges. Many tumors are not detected by these approaches when tested on public datasets for tumor detection and segmentation such as the Kidney Tumor Segmentation Challenge (KiTS) and the Liver tumor segmentation challenge (LiTS). False negative rates by lesion as high as 50% are commonly observed, and this rate is even higher for smaller lesions as they exhibit a high degree of variability (heterogeneity) among themselves. Additionally, in numerous instances, these lesions share similarities (homogeneity) in intensity, size, and shape with other anatomical structures as well as blurriness and blending with surrounding tissue. To improve the detection and segmentation accuracy of lesions in CT scans of the liver and kidney, we propose a selective ensemble approach that uses the predictions of two models to select the best possible mask for lesions. Both models are based on the UNet architecture and use the ConvNext convolutional block in both the encoder and decoder. The first model is trained on lesion segmentation regardless of size, while the second is designed and fine-tuned to segment and detect small lesions. Once the segmentation mask is predicted from both models we extract intensity-based features from within the lesion, contrast them with features from surrounding tissue, and select the mask that maximizes features' separation between the two. We test our approach on three different datasets for lesion segmentation in the kidney and liver. Our proposed approach achieves an improved detection and segmentation performance and is able to increase the number of lesions detected in all three datasets when compared to current state-of-the-art models.