Scholarlink

<p>Pulmonary nodule segmentation and severity assessment are crucial for the diagnosis and monitoring of pulmonary fibrosis, enabling accurate identification of fibrotic changes and evaluation of disease progression. Current architectures frequently face substantial computational demands, limited classification accuracy in both binary and multi-class tasks, and limited interpretability. To mitigate the identified drawbacks, this research presents a Tornado Optimizer with Coriolis force (TOC)-enriched Lightweight Volume-weighted Physics-informed Convolutional Neural Network (TOC-LVPCNN), improving feature learning and classification performance. With the use of physics-informed priors and a lightweight architecture, the model gets better at extracting features and, at the same time, consumes less computational power, which makes it possible to perform quicker inference and to be more efficient. The TOC goes hand in hand with the classification process in taking care of the hyperparameters and the weight updates, making decision boundaries more accurate and training faster. Besides, the model adds SHapley Additive exPlanations (SHAP) technique for interpretability, which allows clinicians to understand how much each feature contributes to the predictions, thus lending transparency, trust, and better validation in the clinicals. Moreover, SHAP not only increases the interpretability of the model but also keeps the model looking at clinically relevant features, thereby diminishing the danger of learning biased patterns stemming from demographic as well as irrelevant features. The TOC-LVPCNN model, which is the one proposed herein, not only delivered outstanding performance but also registered a classification accuracy of 98.5, % precision of 97.8%, a recall of 98.1%, an F1-score of 97.9%, and a Dice coefficient of 96.7% showcasing its efficacy and durability in both the categorization and segmentation tasks. Therefore, this model effectively addresses key challenges in pulmonary nodule analysis by combining physics-informed learning and explainability, achieving high accuracy and robustness in both segmentation and classification tasks for reliable clinical application.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An optimized hybrid deep learning model with SHapley Additive exPlanations for pulmonary fibrosis nodule segmentation and severity assessment

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月21日

查看全文
  • Pallavi Patil,
  • S. Supreeth,
  • Kirankumari Patil

摘要

Pulmonary nodule segmentation and severity assessment are crucial for the diagnosis and monitoring of pulmonary fibrosis, enabling accurate identification of fibrotic changes and evaluation of disease progression. Current architectures frequently face substantial computational demands, limited classification accuracy in both binary and multi-class tasks, and limited interpretability. To mitigate the identified drawbacks, this research presents a Tornado Optimizer with Coriolis force (TOC)-enriched Lightweight Volume-weighted Physics-informed Convolutional Neural Network (TOC-LVPCNN), improving feature learning and classification performance. With the use of physics-informed priors and a lightweight architecture, the model gets better at extracting features and, at the same time, consumes less computational power, which makes it possible to perform quicker inference and to be more efficient. The TOC goes hand in hand with the classification process in taking care of the hyperparameters and the weight updates, making decision boundaries more accurate and training faster. Besides, the model adds SHapley Additive exPlanations (SHAP) technique for interpretability, which allows clinicians to understand how much each feature contributes to the predictions, thus lending transparency, trust, and better validation in the clinicals. Moreover, SHAP not only increases the interpretability of the model but also keeps the model looking at clinically relevant features, thereby diminishing the danger of learning biased patterns stemming from demographic as well as irrelevant features. The TOC-LVPCNN model, which is the one proposed herein, not only delivered outstanding performance but also registered a classification accuracy of 98.5, % precision of 97.8%, a recall of 98.1%, an F1-score of 97.9%, and a Dice coefficient of 96.7% showcasing its efficacy and durability in both the categorization and segmentation tasks. Therefore, this model effectively addresses key challenges in pulmonary nodule analysis by combining physics-informed learning and explainability, achieving high accuracy and robustness in both segmentation and classification tasks for reliable clinical application.