Work

Deep Learning for Lesion Classification in CT Imaging: A Systematic Literature Review

Year: 2026

Type: preprint

Abstract: Accurate lesion classification in computed tomography (CT) imaging is crucial for early diagnosis and effective clinical decision-making. With the growing adoption of low-dose CT (LDCT) to reduce radiation exposure and the rapid evolution of deep learning (DL), automated lesion analysis has gained significant momentum. Despite this progress, existing research often lacks a unified assessment of model design, generalizability, and training practices. This systematic literature review (SLR) analyzes recent DL approaches for lesion classification from CT images, focusing on the influence of model design and training strategies on performance and generalizability. A comprehensive search was conducted across Web of Science, PubMed, Scopus, and IEEE Xplore. Following a rigorous screening and selection process in accordance with PRISMA guidelines, 57 peer-reviewed articles published between January 2020 and December 2025 were included. The findings show that lung and abdominal lesions were the primary research focus, with CNN-based models remaining the dominant approach because of their effectiveness in learning local morphological and textural features. Transformer-based models demonstrated superior contextual learning, while hybrid architectures generally achieved the best performance by integrating complementary local and global feature representations. Self-supervised and transfer learning further improved generalizability in data-limited settings. However, most studies relied on retrospective, small, and class-imbalanced datasets with limited external validation, restricting reproducibility and clinical translation. Persistent challenges include the limited availability of public LDCT datasets, insufficient model interpretability, and inadequate evaluation under real-world clinical conditions. Future research should prioritize large multi-center datasets, robust external validation, explainable AI, and clinically representative evaluation frameworks to develop reliable and generalizable CT-based lesion classification systems.

Source: Research Square

DOI: 10.21203/rs.3.rs-10632411/v1