We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
An Improved Deep Network Model to Isolate Lung Nodules from Histopathological Images Using an Orchestrated and Shifted Window Vision Transformer.
- Authors
Sabitha, Ponnan; Canessane, Ramalingam Aroul; Pillai Minu, Manickarasi Sivathanu; Gowri, Vinayagamoorthy; Antony Vigil, Maria Soosai
- Abstract
Cancer is a major health issue worldwide. Classification of pulmonary (lung) nodules into benign and malicious is one of the stimulating exploration domain as it is the second most serious malignancy and the crucial source of universal deaths. Accurate identification of lung cancer from Computed Tomography (CT) scans achieves an important role in cancer diagnostics system. Besides, the accuracy of the manual isolation framework for lung cancer is dependent on the severity of the malignancy and the efficiency of the radiologist, which frequently cause inappropriate decisions. Thus, the segmentation of the affected area from the CT images is a very challenging task since the morphological features of pulmonary nodules are very complex. Recently, Machine Learning (ML) approaches, particularly Deep Learning (DL) methods enable medical industry to analyse huge data at remarkable speeds without debasing the accuracy of tumour segmentation algorithms. However, due to minute inter-class variances between the affected area and its adjacent tissues and the huge diversity of isolation targets, the deep models often fail to segment lung nodules accurately. To solve these issues, we develop an Orchestrated and Shifted Window Transformer (OSWT) with Multi-head self-attention (MSA) units to isolate the abnormal (diseased) area from pulmonary CT images precisely. We assess OSWT on a CT lung image dataset, called The Cancer Genome Atlas (TCGA or Atlas), and relate the performance of the proposed OSWT against 7 innovative classification models in terms of performance measures. The segment or using an OSWT delivers 98.4% dice similarity index (DSI), 96.5% of Jaccard similarity measure (JSM), 0.73% of volume error (VE), and 0.99s average computational cost. The extensive experimental results demonstrate that the OSWT model realizes improved performance and is more suitable for isolating abnormal cancer area from CT scans.
- Subjects
TRANSFORMER models; PULMONARY nodules; DEEP learning; COMPUTED tomography; MACHINE learning; LUNGS
- Publication
Traitement du Signal, 2024, Vol 41, Issue 4, p2081
- ISSN
0765-0019
- Publication type
Article
- DOI
10.18280/ts.410436