Academic achievements

Peer-reviewed academic journal articles

  • Michel Pohl, Mitsuru Uesaka, Hiroyuki Takahashi, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, "Frame forecasting in cine MRI using the PCA respiratory motion model: comparing recurrent neural networks trained online and transformers", Computerized Medical Imaging and Graphics (2026)
    [arXiv preprint] [Published version] [Code on GitHub] This study applies RNN online learning algorithms for the first time to frame forecasting in chest and liver cine MRI and compares them with transformers. These algorithms are integrated within a modular, interpretable, and data-efficient pipeline combining deformable image registration, PCA-based respiratory motion modeling, and temporal breathing-dynamics prediction. Performance was broadly in line with prior literature on respiratory MRI forecasting, including works exploring end-to-end generative architectures. Online-trained RNNs outperformed other methods at medium-to-long horizons (the time interval in the future for which the prediction is made), while sequence-specific transformers were competitive at low-to-medium horizons.
  • Michel Pohl, Mitsuru Uesaka, Hiroyuki Takahashi, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, "Real-time respiratory motion forecasting with online learning of recurrent neural networks for accurate targeting in externally guided radiotherapy", Computer Methods and Programs in Biomedicine (2025)
    [arXiv preprint] [Published version] [Code on GitHub] This study investigates RNN online learning algorithms, including Decoupled Neural Interfaces (DNI) and Sparse One-Step Approximation (SnAp-1), for respiratory motion forecasting from external chest and abdomen markers, examining the combined effect of sampling rate and horizon on their performance. I exploited sparsity in the influence and immediate Jacobian matrices to derive a compressed formulation of SnAp-1 with lower memory and computational complexity. For DNI, I derived improved coefficient-update equations for credit-assignment estimation, yielding more accurate forecasts. Despite sequence-wise training and testing on a small dataset with irregular breathing records, online-trained RNNs achieved accuracy comparable to prior deep learning approaches trained on substantially larger datasets.
  • Michel Pohl, Mitsuru Uesaka, Hiroyuki Takahashi, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, "Prediction of the position of external markers using a recurrent neural network trained with unbiased online recurrent optimization for safe lung cancer radiotherapy", Computer Methods and Programs in Biomedicine (2022)
    [arXiv preprint] [Published version] [Code on GitHub] [Blog article] This study is the first to apply Unbiased Online Recurrent Optimization (UORO), a computationally efficient online learning algorithm for RNNs that approximates Real-Time Recurrent Learning (RTRL), to respiratory motion forecasting through prediction of external thoracic and abdominal marker positions. I proposed an efficient implementation of UORO for vanilla RNNs by deriving closed-form expressions governing the recursive update of rank-one stochastic estimates of the influence matrix (i.e., the state–parameter Jacobian), from which unbiased online loss-gradient estimates are obtained. UORO yielded the lowest RMSE among the algorithms compared (1.3 mm across horizons below 2.0 s) while remaining robust to irregular breathing patterns. It also maintained prediction times below 2.8 ms per time step, compared with over 55 ms for RTRL (Dell Intel Core i9-9900K 3.60 GHz).
  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, "Prediction of the motion of chest internal points using a recurrent neural network trained with real-time recurrent learning for latency compensation in lung cancer radiotherapy", Computerized Medical Imaging and Graphics (2021)
    [arXiv preprint] [Published version] [Code: Lucas–Kanade optical flow] [Code: RTRL] [Code: Nadaraya–Watson image warping] I developed a method to reconstruct and forecast 3D lung tumor ROIs in time-resolved CT acquisitions from an initial frame and the trajectories of internal points tracked with the Lucas–Kanade optical-flow algorithm. This approach combines a linear correspondence model relating the motion of these points to tissue motion with RNNs trained using RTRL, a classical online learning algorithm for RNNs based on exact recursive computation of the influence matrix. RTRL yielded a maximum prediction error of 1.5 mm for the tracked points, below the 2 mm clinical guideline threshold, and a Pearson correlation coefficient of 0.95 between predicted and ground-truth images. Tumor positions in the predicted frames remained qualitatively consistent with the ground truth.

Technical reports

Conferences in Japan (highlight)

  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, “Prediction of the position of external markers on the chest and abdomen using a recurrent neural network trained with real-time recurrent learning for accurate and safe lung cancer radiotherapy”, 2021 Spring Meeting of the Atomic Energy Society of Japan, online, March 2021
    [Presentation information] [One-page abstract]
  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, and Ritu Bhusal Chhatkuli, "3D internal points position prediction using a recurrent neural network for tumor tracking during lung cancer radiotherapy", 2020 Spring Meeting of the Atomic Energy Society of Japan, Fukushima, March 2020
    [Presentation information] [One-page abstract]
  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, Ritu Bhusal Chhatkuli and Akihiro Haga, "CT slice prediction with optical flow for delay compensation in image-guided lung radiotherapy", 12th Students' research meeting of the Kanto-Koetsu branch of the Atomic Energy Society of Japan, Tokyo, March 2019
  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, Ritu Bhusal Chhatkuli, and Akihiro Haga, "Combining optical flow and principal components analysis for tumor motion analysis during X-ray radiotherapy", 2018 Spring Meeting of the Atomic Energy Society of Japan, Osaka, March 2018
    [Presentation information] [One-page abstract]

International workshops

  • Michel Pohl, Mitsuru Uesaka, Kazuyuki Demachi, Ritu Bhusal Chhatkuli, and Akihiro Haga, "Optical flow for the calculation of lung motion during X-ray radiotherapy", SNU & Utokyo Joint Workshop, Seoul, July 2017

Scholarships, grants, and awards

  • Sept. 2018–Sept. 2021: SEUT-RA (The University of Tokyo Doctoral Student Special Incentives Program). Competitive doctoral research grant; received support throughout 2018–2021, including SEUT-RA Type A in 2019–2020, awarded to the top 20 second-year Ph.D. applicants at the Graduate School of Engineering.
  • Apr. 2019–Apr. 2021: Scholarship from the Epson International Scholarship Foundation. Two-year scholarship awarded following a competitive selection process, including interviews in Japanese at Epson on research activities and future plans.
  • March 2018 and March 2019: Idea Prize and Encouragement Prize, Atomic Energy Society of Japan. Awarded for my poster at the AESJ spring annual meeting and my presentation at the AESJ student meeting, respectively.
  • 2016: Valedictorian award (M.Sc. in Signal and Image Processing, Centrale Méditerranée).
  • Sept. 2016–March 2017 and Sept. 2017–March 2018: MEXT Honors Scholarship for Privately-Financed International Students.

Service

  • 2023: reviewer for IEEE Access