The history of radiotherapy is also the history of the development of medical technology: X-rays were discovered by Dr. Roentgen in 1895, and palliative irradiation using X-rays was implemented the following year. Radiotherapy then progressed from 2D to 3D with the generation of stable high-energy X-rays and advances in diagnostic imaging, especially with the advent of computed tomography (CT), which made it possible to depict tumors not as shadows but as three-dimensional (3D) structures. The improved computational power of computers has also made it possible to irradiate complex geometries. Initially, intensity-modulated radiation therapy (IMRT) could only be applied to tumors whose shape did not easily change, even if a large amount of preparation time was spent on it, due to the calculation time required. IMRT is now used in many clinical situations. In this review, I will discuss the use of artificial intelligence (AI) in the field of radiotherapy, not from a technical perspective, but from a clinician’s viewpoint, with the expectation that AI will improve the quality of life of both patients and medical staff in addition to standardizing treatment.

Explainable artificial intelligence (XAI) is a research field on the techniques which aim to reduce the opacity of high-performance machine learning models, including deep neural networks and ensemble models, and it has been paid much attention especially in medicine. In this article, I first review several early attempts related to XAI in the literature of artificial intelligence. Then, key concepts, representative techniques, and open issues in XAI are discussed in depth. Finally I also present several advances in the recent XAI research.

Endoscopy is a critical method for detecting early-stage gastric cancer; however, its complexity and the potential for oversight remain significant challenges. To address this, we developed an artificial intelligence (AI)-based system to classify endoscopic images using contrastive language-image pre-training (CLIP), a multimodal AI model, to support gastric cancer screening. The dataset collected at Fujita Health University Hospital consists of endoscopic images and corresponding textual descriptions of disease names and conditions. Leveraging this data, we employed multitask learning with the CLIP model to classify gastric lesions. The proposed method achieved an average accuracy of 0.834 for four-class classification and 0.831 for three-class classification, highlighting its effectiveness in supporting gastric cancer diagnosis.

Ablation therapy, a treatment for atrial fibrillation (AF), is known to have a higher recurrence rate as the duration of AF increases. In this study, we used contrast-enhanced CT images to analyze the region including the left atrium in the images to classify paroxysmal AF, which can be treated with early intervention and is expected to be cured, and long-term AF, which has a high recurrence rate, by analyzing the region including the left atrium in the images. We aimed to classify paroxysmal AF, which can be treated with early intervention and is highly treatable, and long-standing, persistent AF, which has a high recurrence rate. We previously classified paroxysmal AF and long-standing persistent AF by inputting contrast-enhanced CT images of 60 patients each with paroxysmal AF and long-standing persistent AF into 6 types of convolutional neural networks. In this study, we attempted to further improve the accuracy of the classification by using contrast-enhanced CT images as input images by cropping only the heart and its surrounding structures, and by focusing only on the surrounding structures of the left atrium. The classification results showed that patients with enlarged left atrium tended to be predicted to have long-lasting AF in many cases, confirming the validity of this method. In addition, many patients in both categories tended to focus on the shape of the left atrium, suggesting that this method can be used classification of AF based on the same standard as that used by physicians. Furthermore, external validation of the method using data collected at other institutions showed that the accuracy of the method was equivalent to that of internal validation, confirming the validity of the method for data from other institutions.