Breast cancer is the main cause of death in women throughout the world. Early detection using ultrasound is very necessary to reduce cases of breast cancer. However, the ultrasound analysis process requires a lot of time and medical personnel because classification is difficult due to noise, complex texture, and subjective assessment. Previous studies were successful in ultrasound classification of breast cancer but required large computations and complex models. This research aims to overcome these shortcomings by using a lighter but more accurate model. We integrated the CBAM attention module into the MobileNetV2 model to improve breast cancer detection accuracy, speed up diagnosis, and reduce computational requirements. Gradient Weighted Class Activation Mapping (Grad-CAM) is used to improve classification explanations. Ultrasound images from two databases were combined to train, validate, and test this model. The test results show that MobileNetV2-CBAM achieves a test accuracy of 93%, higher than the complex models VGG-16 (80%), VGG-19 (82%), InceptionV3 (80%), and ResNet-50 (84%). CBAM is proven to improve MobileNetV2 performance with an 11% increase in accuracy. Grad-CAM visualization shows that MobileNetV2-CBAM can better focus on localizing important regions in breast cancer images, providing clearer explanations and assisting medical personnel in diagnosis.
1 January 2020 in IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences
Numerous applications such as autonomous driving, satellite imagery sensing, and biomedical imaging use computer vision as an important tool for perception tasks. For Intelligent Transportation Systems (ITS), it is required to precisely recognize and locate scenes in sensor data. Semantic segmentation is one of computer vision methods intended to perform such tasks. However, the existing semantic segmentation tasks label each pixel with a single object's class. Recognizing object attributes, e.g., pedestrian orientation, will be more informative and help for a better scene understanding. Thus, we propose a method to perform semantic segmentation with pedestrian attribute recognition simultaneously. We introduce an attribute-aware loss function that can be applied to an arbitrary base model. Furthermore, a re-annotation to the existing Cityscapes dataset enriches the ground-truth labels by annotating the attributes of pedestrian orientation. We implement the proposed method and compare the experimental results with others. The attribute-aware semantic segmentation shows the ability to outperform baseline methods both in the traditional object segmentation task and the expanded attribute detection task.
11 April 2024 in International Journal of Computational Intelligence Systems
As we all know, cancer is one of the leading causes of death worldwide and the second leading cause of death overall. This is why regular screenings or health checks are necessary to detect cancer lesions early. Since bone scan images have become the primary means of detecting the emergence of cancer lesions on bone, high segmentation accuracy is essential for establishing the model of some predefined regions in bone scan images where cancer metastasis was predicted to appear. Consequently, robust localization and identification of the specific region in bone scan images are required for automated metastasis detection. To this end, we propose Efficient-BtrflyNet, a new deep learning-based architecture for skeleton segmentation of whole-body bone scan images. The proposed architecture exploits the benefits of EfficientNet’s model scaling and the encoder–decoder design of butterfly-type networks. We added EfficientNetB7 to the encoder section to obtain more specific features. The proposed architecture simultaneously processes anterior and posterior whole-body bone scan images. Using 37 bone scan images, we evaluated the performance of our proposed skeleton segmentation system using the Dice score. Efficient-BtrflyNet achieves superior segmentation performance compared to the existing representative method.
6 December 2020 in IEEJ Transactions on Electrical and Electronic Engineering
In Balinese culture, the ability to read Balinese script is one of the challenges young generations face. Advances in machine learning have proposed handwriting detection systems using both traditional and deep learning models. However, the traditional approach is usually impractical and is prone to inaccurate identification results. Convolutional neural network (CNN)-based models integrate feature extraction and classification into an end-to-end pipeline to increase performance. This research proposes that recognizing characters through an object detection approach makes an end-to-end process of localizing and classifying several characters simultaneously using the Faster R-CNN. Four CNN models, including ResNet-50, ResNet-101, ResNet-152, and Inception ResNet V2, were tested to detect 28 Balinese characters in a single form that covers 18 consonants and 10 digits using Intersection over Union (IoU) thresholds: 0.5 and 0.75. ResNet-50 and Inception ResNet V2 achieve 0.991 mAP at IoU of 0.5, while Inception ResNet V2 excels at IoU of 0.75. Further analysis showed that the class ‘nol’ had the lowest Recall due to many undetected ground truths. Meanwhile, class ‘ba’ had the lowest Precision due to its similarity to classes “ga” and “nga”. This research contributes to the experiment with Faster R-CNN in detecting handwritten Balinese scripts.
27 September 2023 in Building of Informatics, Technology and Science (BITS)
Colorectal Cancer (CRC) is a major contributor to cancer-related mortality worldwide, necessitating early detection and treatment of polyps to prevent cancer progression. A colonoscopy is a critical diagnostic procedure for identifying colon abnormalities and removing premalignant polyps. However, accurately segmenting polyps in colonoscopy images poses challenges due to their diverse appearance and indistinct boundaries. In this study, we investigate augmentation techniques to enhance polyp semantic segmentation using the U-Net model. Our analysis reveals that the most effective technique is found in sub-scenario 2.6.c with an input size of 320×320, striking a favorable balance between accuracy and efficiency. Additionally, we explore the benefits of larger input sizes, taking into account resource considerations. Moreover, we conduct further testing of the best augmentation technique identified in previous experiments with the SegNet model. The results show a 3.5% improvement in the dice coefficient and slightly better qualitative outcomes. However, it is important to note that this enhancement comes with a nearly fivefold increase in training time. Moving forward, our objective is to develop a unified model for segmenting diverse medical images, pushing the boundaries of polyp detection and medical imaging. This research provides valuable insights and lays the foundation for more advanced applications in polyp detection and medical image analysis.