Research
My goal is to support human decision-making with image processing and pattern recognition. Computer-aided diagnosis of medical images is the core, and the subjects have also extended beyond medicine.
Medical Image Analysis
Ultrasound
Detection and tracking of liver tumors in abdominal ultrasound videos. We aim at detection that works in real time during examination and stays consistent across frames. We also address how to learn from training data with incomplete annotations.
X-ray CT
Detection of small lung cancers and measurement of treatment response, classification of pulmonary arteries and veins, liver segmentation and liver cancer detection, and registration of time-series CT images. Quantifying the convergence of vessels and bronchi around lesions with a 3-D concentration index has been a long-standing theme.
Endoscopy and laparoscopy
Intraoperative support, including surgical instrument segmentation using depth estimated from monocular laparoscopic video, and surgical navigation triggered by spoken anatomical names.
Character Recognition, Language and Historical Documents
Modern official documents and historical materials
In character recognition of modern Japanese official documents, a key issue is how to handle imbalanced training data. Since recognition results still contain errors, we now use large language models (LLMs) to correct misreadings based on context and document-specific vocabulary.
Generating gikun readings based on sound symbolism
Gikun is a Japanese writing practice in which kanji are given readings that match their meaning or feeling rather than their standard readings. We model the impressions carried by sounds (sound symbolism) with a Bi-LSTM and combine them with word meanings captured by document embeddings to generate natural gikun automatically.
Handwritten characters
Active collection of character data to improve recognition accuracy, which leads to the question of what data allows a classifier to learn efficiently.
Understanding Human Activity
Sports
Form analysis of sprinters based on joint positions, alignment between videos, and measurement of instantaneous swimming speed with a hand-held camera.
Food and daily life
Analysis and estimation of the factors that make food photographs attractive.
Past Research
- Detection of variable star candidates from time-series astronomical images (narrowing down with anomaly detection)
- A virtual endoscopy environment for the stomach
- A virtual needle biopsy simulator for the prostate
- Medical image processing on FPGAs (with the Yokota Laboratory, Utsunomiya University)
- Benign/malignant classification of thoracic tumors by quantifying the spatial relationship between vessels and tumors
- Lesion detection in double-contrast X-ray images of the stomach
- A speech-recognition-based communication system for people with physical disabilities
Results are listed on the Publications page. For the laboratory's activities, see MDLab (in Japanese).