Approaching Cross-Disease Features for Improved Classification of Thyroid and Breast Cancer in Ultrasound Images
Mohamed Bal-Ghaoui, My Hachem El Yousfi Alaoui, Abdelilah Jilbab, Abdennacer Bourouhou · 2023
Thyroid and breast cancers are seen astwo diseases that mostly affect women worldwide. Ultrasonography is known to bean effective modality in interpreting these diseases. It is a safe, accessible,cost-effective, and non-invasive but highly operator-dependent technique.Ultrasound (US) images suffer from poor image quality. Thus, the need for anexperienced radiologist for an effective assessment of these images. Computer-Aided-Diagnosis (CAD) systems based on artificial intelligence can reliablyease the interpretation of US images and provide assistance to radiologists. Inthis article, we propose a previously validated Convolutional Neural Network(CNN) architecture for breast ultrasound to train and classify thyroid noduleswhile assessing the effectiveness of this approach by leveraging the shared USfeatures between the two diseases. We also evaluated five state-of-the-artpre-trained models on both breast and thyroid nodules using the sameclassifier. Our customized CNN achieved 94.84% and 85.35% accuracy whilemaintaining a low positive rate of 5.5% and 13.0% on both open-access datasetsfor breast and thyroid, respectively. Regarding Transfer Learning, InceptionV3and VGG16 are seen to outperform other models.