Interpretable features fusion with precision MRI images deep hashing for brain tumor detection. (April 2023)

• Record Type: Journal Article
• Authors: Özbay, Erdal; Altunbey Özbay, Feyza
• Is Part Of: Computer methods and programs in biomedicine. Volume 231(2023)
• Publication Date: 2023-04
• Subjects: Brain tumor -- Deep hashing -- Feature fusion -- Image retrieval -- Interpretability; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2023.107387
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: The high-class similarity in Brain-Tumor-MRI (BT-MRI) images and masking of relatively smaller tumor regions to increase the recognition rate and retrieval accuracy of hash codes, highlight key finding features with an interpreted saliency mapping mechanism. In order to solve the trouble of information loss due to the noises contained in traditional MRI, the accuracy results of full information representation retrieval are increased by using the finding features that can prevent tumor omissions by preserving semantics from two separate networks, local and global. Poor ranking performance and retrieval accuracy are increased with bit-balanced, cross-entropy, and quantization losses used to fine-tune quality performance and ranking quality in the retrieval process of medical images. Abstract: Background and Objective: Brain tumor is a deadly disease that can affect people of all ages. Radiologists play a critical role in the early diagnosis and treatment of the 14, 000 persons diagnosed with brain tumors on average each year. The best method for tumor detection with computer-aided diagnosis systems (CADs) is Magnetic Resonance Imaging (MRI). However, manual evaluation using conventional approaches may result in a number of inaccuracies due to the complicated tissue properties of a large number of images. Therefore a precision medical image hashing approach is proposed that combines interpretability and feature fusion using MRI images of brain tumors, to address the issue of medical image retrieval. Methods: A precision hashing method combining interpretability and feature fusion is proposed to recover the problem of low image resolutions in brain tumor detection on the Brain-Tumor-MRI (BT-MRI) dataset. First, the dataset is pre-trained with the DenseNet201 network using the Comparison-to-Learn method. Then, a global network is created that generates the salience map to yield a mask crop with local region discrimination. Finally, the local network features inputs and public features expressing the local discriminant regions are concatenated for the pooling layer. A hash layer is added between the fully connected layer and the classification layer of the backbone network to generate high-quality hash codes. The final result is obtained by calculating the hash codes with the similarity metric. Results: Experimental results with the BT-MRI dataset showed that the proposed method can effectively identify tumor regions and more accurate hash codes can be generated by using the three loss functions in feature fusion. It has been demonstrated that the accuracy of medical image retrieval is effectively increased when our method is compared with existing image retrieval approaches. Conclusions: Our method has demonstrated that the accuracy of medical image retrieval can be effectively increased and potentially applied to CADs.