LMLCC-Net: A Semi-Supervised Deep Learning Model for Lung Nodule Malignancy Prediction from CT Scans using a Novel Hounsfield Unit-Based Intensity Filtering

Tasnia Binte Mamun, Adhora Madhuri, Nusaiba Sobir, Taufiq Hasan · 2025

Read LMLCC-Net: A Semi-Supervised Deep Learning Model for Lung Nodule Malignancy Prediction from CT Scans using a Novel Hounsfield Unit-Based Intensity Filtering by Tasnia Binte Mamun, Adhora Madhuri, Nusaiba Sobir, Taufiq Hasan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06370.txt.

Introduction

Tasnia Binte Mamun and co-authors teach a network to judge whether a lung nodule is cancerous, using a clue earlier systems ignored. Their 2025 paper works from CT scans, where benign and malignant nodules differ measurably in their brightness profile on the Hounsfield scale. The model filters the image by that intensity before judging shape and texture.

The design is semi-supervised, learning partly from unlabeled scans, and reads the nodule in three dimensions through several branches, each with its own learned intensity filter. Lung cancer kills more patients than any other, so catching nodules early and sorting them right carries real weight. The intensity-first trick is the paper's main contribution.

Catalog record

CREATORTasnia Binte Mamun, Adhora Madhuri, Nusaiba Sobir, Taufiq Hasan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-06370

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LMLCC-Net: A Semi-Supervised Deep Learning Model for Lung Nodule Malignancy Prediction from CT Scans using a Novel Hounsfield Unit-Based Intensity Filtering Tasnia Binte Mamun, Adhora Madhuri, Nusaiba Sobir, Taufiq Hasan Abstract Lung cancer remains one of the leading causes of mortality worldwide. Early diagnosis of malignant pulmonary nodules in CT images can significantly reduce disease-related mortality and morbidity. This study aims to develop a novel deep learning framework that leverages the unique intensity characteristics of nodules, measured in Hounsfield Units, for improved classification of benign and malignant pulmonary nodules. In this work, we propose LMLCC-Net, a novel deep learning framework for classifying nodules from CT scan images using a three-dimensional convolutional neural network, incorporating Hounsfield Unit based intensity filtering. Benign and malignant nodules exhibit significant differences in their Hounsfield Unit based intensity profiles, which have not been sufficiently explored in the existing literatures. Our method considers both the intensity pattern and the texture for malignancy prediction. LMLCC-Net extracts features from multiple branches, each using a separate, learnable Hounsfield Unit based intensity filtering stage. Various combinations of branches and learnable filter ranges were explored to identify the best-performing model. In addi tion, we propose a semi-supervised learning scheme for labeling ambiguous cases and develop a lightweight model to efficiently classify the nodules. The proposed LMLCC-Net was evaluated using the LUNA16 dataset. Our proposed method achieves a classification accuracy of 91.96%, a sensitivity of 92.94%, and an area under the curve of 94.07%, showing improved performance compared to existing methods. Our proposed LMLCC-Net framework, which effectively combines Hounsfield Unit-based intensity filtering and texture analysis, significantly improves the classification performance of pulmonary nodules in CT images. By utilizing a multi branch architecture and semi-supervised learning, the model enhances its ability to handle ambiguous cases while maintaining computational efficiency through a lightweight design. † † history: Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000. † † doi: 10.1109/ACCESS.2024.0429000 † † address: mHealth Lab, Department of Biomedical Engineering, Bangladesh University of Engineering and Technology (BUET), Dhaka-1205, Bangladesh. † † titlenote: ‡ These authors contributed equally. Correspondence to: Taufiq Hasan (email: [email protected]). † † corresponding: Corresponding author: Taufiq Hasan (e-mail: [email protected]). I Introduction Lung cancer is the leading cause of cancer deaths worldwide. It was the most frequently diagnosed cancer in 2022, responsible for almost 2.5 million new cases, or one in eight cancers worldwide (12.4% of all cancers globally) [ 1 ] . Early detection is crucial to improving the survival rates of lung cancer patients, as it allows for effective and timely treatment interventions [ 2 ] . Low-dose computed tomography (CT) screening has emerged as a powerful tool for the early detection of lung cancer, enabling the identification of lung nodules that may indicate malignancy [ 3 ] . However, the implementation of annual low-dose CT screenings substantially increases radiologists’ workload, particularly in developing countries where there is a pronounced shortage of expert and qualified radiologists [ 4 ] . Another challenge is that lung nodules, which can be solid, partially solid, or ground glass, each with varying cancer probabilities, exhibit extraordinary spatial and temporal heterogeneity at multiple levels, including genes, proteins, cells, microenvironment, tissues, and organs [ 5 ] , leading to missed diagnoses and reporting errors. This underscores the critical need for efficient, accurate and automated methods to assist radiologists in the detection and prediction of lung nodule malignancy. Most of the previous studies on convolutional neural network (CNN) based lung nodule analysis methods [ 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 ] in this area have used the publicly available Lung Image Database Consortium image collection (LIDC-IDRI) dataset [ 23 ] . These studies calculated the mean malignancy level (MeanML) for each nodule. However, this labeling method has limitations in accuracy. A significant amount of data with potential labeling inaccuracies can adversely affect model training. In addition, no prior study has explored the distribution patterns of HU within lung nodules or considered focusing on distinct HU ranges as a feature for classification. Benign tumors typically exhibit smooth, regular borders, while malignant tumors often present with irregular borders [ 18 , 19 , 20 ] . Additionally, malignant tumors have been shown to possess higher values of vascularity index and maximal shear velocity, indicating that malignant lesions tend to be stiffer compared to benign or intermediate tumors [ 21 ] . These characteristics influence the HU distribution within nodules across different ranges, making it a valuable feature for distinguishing between benign and malignant tumors. Example distributions of HU intensity profiles in benign and malignant tumors are represented in Fig. 2 . Furthermore, to better visualize how the Hounsfield Unit-based segmentation captures intensity-specific regions within a nodule, representative images were generated by segmenting the HU ranges into multiple sub-intervals. As shown in Fig. 1 , this process isolates intensity-dependent structural variations that correspond to different tissue densities, allowing the model to focus on diagnostically relevant sub-regions such as soft tissue, fibrotic cores, and calcified components. The main contributions of our study can be summarized as follows. 1) We develop a novel approach to labeling the LUNA16 dataset, implementing a semi-supervised method specifically designed to handle ambiguous cases. This strategy enhances the accuracy of the labels, resulting in more reliable data for model training. 2) We propose the Learnable Multi-branched Lung Cancer Classification Network (LMLCC-Net), which processes CT volumes through multiple HU-based branches to extract complementary texture and density information for malignancy prediction. We intoduce a Learnable Dynamic Range Layer within each branch that automatically optimizes HU intensity boundaries during training, enabling adaptive feature extraction and enhancing generalization across heterogeneous nodules. 3) We conduct extensive validation across multiple architectural configurations including different multibranch setups, learnable versus fixed HU ranges, and constant versus random initialization strategies to confirm the stability and generalization of the proposed framework. Grad-CAM–based interpretability analysis further verifies that the network consistently attends to diagnostically relevant nodule regions. This systematic evaluation ensures the model’s robustness and reliability, making it well-suited for deployment in real-world digital imaging and clinical screening environments. Fig. 1: Visualization of HU range segmentation for a representative nodule. (a) HU range divided between 0–0.3, highlighting low-intensity regions; (b) 0.3–0.7, capturing mid-range densities; and (c) 0.8–1.0, emphasizing high-density structures. This division helps isolate subtle tissue-specific variations that support malignancy prediction. Fig. 2: Intensity range corresponding to the Hounsfield Unit (HU) distribution: (a-d) benign nodules and (e-h) malignant nodules. The remainder of this paper is structured as follows: Section II includes a brief overview of the related works. Section III describes the dataset. Section IV provides a detailed description of the proposed model. Section V features a comprehensive overview of the training parameters and dataset splitting. Section VI presents our results and analysis. Finally, Section VII covers related issues and concludes with our findings. II Related Works Recent advancements in conventional computer-aided diagnosis (CAD) systems have shown significant promise in automating the detection and classification of pulmonary nodules. Conventional computer-aided diagnosis (CAD) systems rely on extracting handcrafted features from lung nodules to build radiomics models for automatic classification [ 6 , 7 , 8 ] . In this approach, various feature sets have

Read the complete text