Department of Computer Science and Engineering, Dhaka University of Engineering & Technology, Gazipur, Gazipur, 1707, Bangladesh.
Department of Computer Science and Engineering, Dhaka University of Engineering & Technology, Gazipur, Gazipur, 1707, Bangladesh.
Comput Biol Med. 2024 Dec;183:109316. doi: 10.1016/j.compbiomed.2024.109316. Epub 2024 Nov 2.
The microwave brain imaging (MBI) system is an emerging technology used to detect brain tumors in their early stages. Multi-class microwave-based brain tumor (MBT) identification and classification are crucial due to the tumor's patterns and shape. Manual identification and categorization of the tumors from the images by physicians is a challenging task and consumes more time. Recently, to overcome these issues, the deep transfer learning (DTL) technique has been used to classify brain tumors efficiently. This paper proposes a Fine-tuned Feature Extracted Deep Transfer Learning Model called FT-FEDTL for multi-class MBT classification purposes. The main objective of this work is to suggest a better pathway for brain tumor diagnosis by designing an efficient DTL model that automatically identifies and categorizes the MBT images. The InceptionV3 architecture is utilized as a base for feature extraction in the proposed FT-FEDTL model. Thereafter, a fine-tuning method is applied to the additional five layers with hyperparameters. The fine-tuned layers are attached to the base model to enhance classification performance. The MBT data are collected from two sources and balanced by augmentation techniques to create a total of 4200 balanced datasets. Later, 80 % images are used for training, 20 % images are utilized for validation, and 80 samples of each class are used for testing the FT-FEDTL model for classifying tumors into six classes. We evaluated and compared the FT-FEDTL model with the three traditional non-CNN and seven pretrained models by applying an imbalanced and balanced dataset. The proposed model showed superior classification performance compared to other models for the balanced dataset. It attained an overall accuracy, recall, precision, specificity, and Fscore of 99.65 %, 99.16 %, 99.48 %, 99.10 %, and 99.23 %, respectively. The experimental outcomes ensure that the proposed model can be employed in biomedical applications to assist radiologists for multi-class MBT image classification purposes. The Anaconda distribution platform with Python 3.7 on the Windows 11 OS is used to implement the models.
微波脑成像 (MBI) 系统是一种新兴技术，用于检测早期脑瘤。由于肿瘤的模式和形状，多类基于微波的脑肿瘤 (MBT) 识别和分类至关重要。医生从图像中手动识别和分类肿瘤是一项具有挑战性的任务，并且需要更多的时间。最近，为了克服这些问题，深度迁移学习 (DTL) 技术已被用于有效地对脑肿瘤进行分类。本文提出了一种名为 FT-FEDTL 的微调特征提取深度迁移学习模型，用于多类 MBT 分类。本工作的主要目的是通过设计一种自动识别和分类 MBT 图像的高效 DTL 模型，为脑肿瘤诊断提供更好的途径。所提出的 FT-FEDTL 模型使用 InceptionV3 架构作为特征提取的基础。此后，应用一种微调方法对附加的五层进行微调，并设置超参数。微调层连接到基础模型以增强分类性能。MBT 数据来自两个来源，并通过扩充技术进行平衡，共创建 4200 个平衡数据集。然后，使用 80%的图像进行训练，20%的图像用于验证，使用每个类的 80 个样本对 FT-FEDTL 模型进行测试，以将肿瘤分类为六个类别。我们通过应用不平衡和平衡数据集，将 FT-FEDTL 模型与三个传统的非 CNN 和七个预训练模型进行了评估和比较。与其他模型相比，该模型在平衡数据集上表现出了优越的分类性能。它在不平衡数据集上分别达到了 99.65%、99.16%、99.48%、99.10%和 99.23%的整体准确率、召回率、精度、特异性和 F 分数。实验结果确保了所提出的模型可以应用于生物医学应用，以协助放射科医生进行多类 MBT 图像分类。在 Windows 11 操作系统上的 Anaconda 发行版平台上，使用 Python 3.7 实现模型。
