Purpose: Automatic detection of liver lesions in CT scans poses a challenge for researchers. The purpose of this study was to evaluate a new deep learning approach that models the variability between metastases boundaries and interiors, and assess whether it can support an automated lesion detection system.
Materials and methods: This research was supported by the ISRAEL SCIENCE FOUNDATION (grant No. 1918/16). Convolutional neural networks (CNN), which are loosely based on the theory of biological neural networks, operates by propagating (“convoluting”) small matrices across the image to create a new filtered layer, this process is repeated to create many layers (deep learning), each layer is built on the previous one. In the end, a loss function provides probabilities from the last layer. A process of “back-propagation” optimizes the network using new images. In this study a multi-class CNN is proposed to categorize input image patches into sub-categories of boundary and interior patches, the decisions of which are fused to reach a binary lesion versus non-lesion decision. Consecutive portal CT examinations were collected and a senior radiologist with 8 years of experience marked the liver and metastases boundaries. True positive and false positive rates of lesion detection by the algorithm were evaluated per case for the entire data set and for lesions > 10 mm.
Results: For validation of our system, we used CT images of 132 livers with 498 2D marked liver metastases. True positive rate and false positive per case were 85.9% and 1.9 for the entire data set and 93.0% and 1.5 for lesions > 10 mm.
Conclusion: Our new deep learning approach algorithm shows promising results in liver metastases detection task. Using prior knowledge of medical data, such as differences between metastases interior and boundaries, may enhance CNN results.
