Hi! I'm Shunsuke Kikuchi, a research ML engineer at Jmees Inc, and a project-researcher at National Cancer Center Japan East Hospital. I'm interested in Machine Learning, surgical naviagation, precision medicine and medical robotics.
Originally from Tokyo-metoropolitan area, I have been studying in the US since 2021. Aiming to be a researcher in the field of AI in medicine.
In my free time, I enjoy jogging, manga/anime, and traveling.
Also as a Kaggle enthusiast, I have participated in various competitions and challenge workshops. Usually choose to work on video-related competition,
but I'm also curious about other fields, including 3D reconstruction, segmentation, and reinforcement learning.
Explore my work, research, and achievements below!
In total hysterectomy, a uterine manipulator is used to stabilize the uterus and ensure safe resection. However, during robot-assisted surgery, intraoperative manipulation of the device cannot be controlled directly by the surgeon, highlighting the need for automation. In this research, we aim to develop a surgical phase classification model that recognizes the manipulator's angle from endoscopic video. The annotation scheme is being refined through close collaboration and repeated discussion between clinicians and annotators to ensure it is suitable for deep learning models.
July 2025 – Present. Working as a project researcher.
SCING is a machine learning model that predicts gene regulatory networks from single-cell & spatial transcriptomics data. Using GRNs from SCING, we are developing scGRNdb, an analysis pipeline with a database of gene regulatory networks for single-cell and spatial transcriptomics data. I was involved in the part of designing community detection algorithm, pipeline development, and Evaluations.
23 Winter - 25 Spring. Working as an Undergraduate Research Assistant.
TMAM a novel framework that adapts any existing 2D segmentation model for video processing by transferring SAM2's memory encoder and attention modules. TMAM applies a memory encoder to past-frame predictions and uses memory attention to refine current-frame features. By leveraging the inherent temporal redundancy in video sequences, TMAM captures contextual cues that may be overlooked by single-frame processing, thereby improving robustness to occlusions and boundary artifacts.
24 Winter - 25 Spring. Leading the project as the first author of the paper.
Read MoreThis research proposes to investigate the efficacy of RgIA4, a selective α9α10 nicotinic acetylcholine receptor (nAChR) antagonist, in alleviating migraine-related pain using rodent models. The study will evaluate RgIA4's dose- and time-dependent effects in rat and mouse models by inducing migraines through environmental stress and inflammatory mediators. The research aims to determine whether blocking α9α10 nAChR can reduce pain and inflammatory responses, potentially offering a new therapeutic approach for migraines with fewer side effects and a reduced risk of medication overuse headaches.
21 Fall - 23 Summer. Working as an Undergraduate Research Scholor.
Silver Medal (27/1950) (Leaderboard)
Compete with machine learning model which predict small molecule-protein interaction using the Big Encoded Library for Chemical Assessment (BELKA)
See more details of our solution here.Bronze Medal (245/2767) (Leaderboard)
Compete with machine learning mode which classify seizures and other partterns of harmful brain activity in critically ill patients.
Task A (CVS Classification): 2nd Prize ($1,850) | Task B (Computationally Efficient, CPU-limited): 3rd Prize ($1,000)
This Lighthouse challenge was selected for its quality and scale within the MICCAI ecosystem.
Segmentation Challenge for Whole Brain Vessel Anatomy — CTA: 3rd, MTA: 3rd
GRS: 2nd | OSATS: 1st | TRACK: 1st
Overall objective: automatic assessment of suturing surgical skills.
3rd
Primary challenge: building robust models for extremely rare positive cases with subtle visual cues in real-world settings.
Task 1 & 2: 1st Prize (Task1, 900€) and 2nd Prize (Task2, 400€)
4-class segmentation challenge for colorectal cancer screening using colonoscopy frames. Developed models to segment anatomy edges and instrument masks.
Task 1 & 2: 3rd Prize (Task1) & 3rd Prize (Task2)
Classification of surgical suturing skills using simulated environment videos.
For the bost tasks, perform surgical tools segmentation and detect tips of then, then put the movements of them into 1DCNN+GRU model to extract features. Final prediction is performed by GBDT.
Task 1 & 2: 3rd Prize(2D-tracking, $675) & 2nd Prize(3D-tracking, $450)
Predicting trajectories in surgical tatoos on tissues in robotic surgery video using the STIR (Surgical Tattoos in Infrared) Dataset.
4th prize (10000 JPY) (Leaderboard)
The task was Semantic segmentation of Mutli-Organ, using one of the largest dataset, TotalSegmentator. Segmented organs include gallbladder, liver, pancreas, spleen, kidneys, adrenal glands, aorta, stomach, and duodenum.
For the best model, we employed 3D-UNet approached with large backbone architecutre SeresNext50x4d, with UNet++ and scse block. Achieved >0.89 in mean DSC. Experimental Source Codes and solutions are here
Jmees Inc. | June - September 2025