Automatic liver and tumor segmentation from computed tomography images using deep learning techniques

According to 2020 statistics of the Global Cancer Observatory (GLOBOCAN), liver cancer is the sixth most frequently detected type of cancer and ranks third among cancer-related deaths. Despite this, detection of liver cancer and lesions is often done manually by radiologists. This manual process is time-consuming, costly, and prone to errors due to the subjectivity of human evaluations. Automating the detection of liver cancer and lesions using images from medical imaging devices such as CT scans, MRIs, and ultrasound is an important area of research in computer vision. Implementation of such automated systems with low error rates can eliminate inconsistencies between experts by standardizing medical image analysis. Thus, it allows faster diagnosis and treatment, preventing disease progression and increasing patient recovery rates. Deep learning, a subset of machine learning, has demonstrated remarkable success in medical imaging, particularly in detecting and classifying the locations and boundaries of organs and lesions. Convolutional neural networks (CNNs), a popular deep learning architecture, have achieved high accuracy in medical image segmentation and classification. Models such as U-Net, a derivative of CNNs, have further developed these capabilities. These achievements highlight the potential of deep learning algorithms to solve significant challenges in medical image processing. Recent developments include models such as DeepLabv3+, which excel at segmenting complex images by capturing details at multiple scales, providing accurate boundaries between different regions in the image. DeepLabv3+ uses extended convolutions that help preserve the resolution of features, making it well-suited for medical imaging tasks where preserving detail is critical for accurate analysis. In this study, a detailed review of medical image segmentation algorithms with deep learning methods is presented. Data preprocessing issues are addressed before data set training processes. Finally, using LiTS, a public dataset, ResNet50, ResNet101, ResNeXt101_32x4d and Efficientnet-b5 models were applied as backbone networks with DeepLabv3 segmentation architecture on abdominal CT images and the results were analyzed. While the highest dice score in liver segmentation was obtained with the ResNet101 model with 0.9674, the highest dice score in tumor segmentation was obtained with the ResNeXt101_32x4d model with 0.7584. The novelty of this work is the use of hybrid algorithms. Hybrid algorithms enable higher accuracy and efficiency to be achieved by combining different models and techniques. These approaches allow more sensitive and accurate detection of liver tumors. The application of hybrid algorithms is providing significant advances in the field of medical imaging and cancer detection, improving patient care and treatment outcomes.