Abstract
Dysplastic cells are vital for diagnosing Subtypes of Lung Adenocarcinoma (SLA) but are difficult to classify due to high morphological similarity. Existing models often suffer from a feature coherency problem, failing to effectively align local dysplastic details with global subtype representations. To address this challenge, we propose Mani-NENet, a manifold neuron energy network-enhancing dysplastic cells perception built on three core innovations. First, we design a Dysplastic Perception Block (DPB) with a novel neuron energy function that prioritizes highly discriminative cell populations, inspired by the similarity between neuronal activation mechanisms and the diagnostic process. Second, we introduce a Differentiated Representation Learning Block (DRLB) in manifold space to amplify subtype-specific features and mitigate the challenge of high intersubtype similarity. Third, we propose a Key Patch Block (KPB) to enhance the extraction of low-level image features, capturing fine-grained morphological details of dysplastic cells. Extensive experiments on the AMU-SLA and AMU-BC datasets demonstrate that Mani-NENet outperforms other 19 state-of-the-art methods, achieving top-ranking average accuracies of 89.21% and 94.71%, respectively. The ablation results demonstrate that our classification performance improves by 1.91%, 2.42%, and 5.54% after applying DPB, DRLB, and KPB, respectively. In addition, we analyze the classification performance of each subclass and achieve accuracies of 94.64%, 96.82%, 97.49%, and 98.66% for LIA, AAH, MPIA, and PIA, respectively. Qualitative analysis shows accurate localization of dysplastic regions, demonstrating strong clinical interpretability and potential for human–machine collaboration.
Funding Information
  • Natural Science Foundation of Chongqing, China, (CSTB2025NSCQ-LZX0012)
  • Beijing Natural Science Foundation (L242114)