Kibrom Berihu Girum

Kibrom Berihu Girum

Researcher in deep/machine learning for medical image analysis

Posts by Collection

portfolio

publications

Deep generative model-driven multimodal prostate segmentation in radiotherapy

Published in Workshop on Artificial Intelligence in Radiation Therapy, MICCAI, 2019

A generative model is proposed to model the modality invariante shape of the prostate.

Recommended citation: Girum KB, Crehange G, Hussain R, Walker PM, Lalande A. Deep generative model-driven multimodal prostate segmentation in radiotherapy. In: Workshop on Artificial Intelligence in Radiation Therapy. Springer, 2019; 119–27.
Download Paper | Download Slides

Fast interactive medical image segmentation with weakly supervised deep learning method

Published in International Journal of Computer Assisted Radiology and Surgery, 2020

This paper proposes a fast, interactive, and weakly supervised deep learning framework for medical image segmentation. Combining a generative network for prior prediction with a CNN for refinement, the method achieved Dice scores up to 97% and processed 3D exams in 0.5 s. It demonstrated strong cross-modality adaptability and suitability for real-time image-guided interventions.

Recommended citation: Girum KB, Créhange G, Hussain R, Lalande A. Fast interactive medical image segmentation with weakly supervised deep learning method. Int J Comput Assist Radiol Surg. 2020; 15: 1437-1444.
Download Paper | Download Slides

Learning with context feedback loop for robust medical image segmentation

Published in IEEE transactions on medical imaging, 2021

This paper introduces a novel deep learning framework, Learning with Context Feedback Loop (LFB-Net), for robust and accurate medical image segmentation. The method addresses challenges in capturing contextual information and inter-region relationships in medical images by integrating a feedback loop into the encoder-decoder architecture. The feedback loop allows the network to revisit and refine its segmentation predictions, improving accuracy and plausibility over time. Echocardiographic image segmentation; Cardiac cine-MRI segmentation, Inner ear segmentation, Prostate segmentation in radiotherapy

Recommended citation: K. Girum, G. Créhange, and A. Lalande, “Learning with context feedback loop for robust medical image segmentation,” IEEE Trans. Med. Imag., vol. 40, no. 6, pp. 1542–1554, Jun. 2021.
Download Paper | Download Slides

Deep learning methods for automatic evaluation of delayed enhancement-MRI. The results of the EMIDEC challenge

Published in Medical Image Analysis, 2022

This paper reports the EMIDEC challenge on using deep learning for myocardial infarction assessment from DE-MRI. Automatic classification of exams as infarct or non-infarct achieved up to 92% accuracy, while myocardium segmentation was effective but struggled in diseased areas. The results show AI’s potential to improve cardiovascular MRI analysis, though further refinement is needed for clinical use.

Recommended citation: Lalande A. et al. (2022).Deep learning methods for automatic evaluation of delayed enhancement-MRI. The results of the EMIDEC challenge. Med. Image Anal., 79 (2022)
Download Paper | Download Slides

Simplicity is All You Need: Out-of-the-Box nnUNet followed by Binary-Weighted Radiomic Model for Segmentation and Outcome Prediction in Head and Neck PET/CT

Published in Head and Neck Tumor Segmentation and Outcome Prediction, MICCAI, 2022

Proposed a binary weighted model for classificaiton and applied nnUNet model for segmentation.

Recommended citation: Rebaud, L., Escobar, T., Khalid, F., Girum, K., Buvat, I.: Simplicity is all you need: out-of-the-box nnUNet followed by binary-weighted radiomic model for segmentation and outcome prediction in head and neck PET/CT. In: Lecture Notes in Computer Science (LNCS) Challenges (2023)
Download Paper | Download Slides

18F-FDG PET maximum-intensity projections and artificial intelligence: a win-win combination to easily measure prognostic biomarkers in DLBCL patients

Published in Journal of Nuclear Medicine, 2022

AI was used to derive surrogate prognostic biomarkers (sTMTV, sDmax) from 2D PET MIP images in DLBCL. A CNN accurately segmented lesions, showing strong correlation with 3D PET/CT biomarkers. The AI-derived biomarkers predicted PFS and OS comparably to traditional methods while reducing computation time, demonstrating potential to streamline clinical workflows.

Recommended citation: Girum, K. B. et al. 18F-FDG P. E. T. Maximum-intensity projections and artificial intelligence: A Win-Win combination to easily measure prognostic biomarkers in DLBCL patients. J. Nucl. Med. 63, 1925–1932 (2022).
Download Paper | Download Slides

Tumor location relative to the spleen is a prognostic factor in lymphoma patients: a demonstration from the REMARC trial

Published in Journal of Nuclear Medicine, 2024

This study introduces three PET/CT biomarkers—SpreadSpleen, Dspleen, and sDspleen—that quantify tumor spread relative to the spleen in DLBCL patients. These biomarkers predict progression-free and overall survival and improve the prognostic value of metrics like TMTV and age-adjusted IPI. Using AI-based spleen segmentation, the features are simple, interpretable, and potentially useful for clinical integration.

Recommended citation: Girum KB, Cottereau AS, Vercellino L, et al. Tumor location relative to the spleen is a prognostic factor in lymphoma patients: a demonstration from the REMARC trial. J Nucl Med 2024;65(2):313–319.
Download Paper | Download Slides

talks

teaching

Assistance lecturer, Electrical and Computer Egnineering

Undergraduate course, Woldia University, Electrical and Computer Engineering Department, 2013

Main activities

Within Electrical and Computer Engineering departments, I prepared comprehensive course materials, updated curricula to align with modern technologies, developed project proposals for student and departmental initiatives, and monitored student performance through continuous assessment and feedback. I facilitated laboratory sessions, supervised capstone and research projects, collaborated with faculty on academic planning and curriculum improvement, integrated industry-relevant engineering tools into coursework, and contributed to academic documentation such as course portfolios, grading rubrics, and quality improvement reports.

Assistance lecturer

Postgraduate Course, Universite de Bourgogne, Health AI Department, 2019

  • Taught deep learning courses to M2 Health AI students at Universite de Bourgogne.
  • Introduction to Convolutional neural networks; introduction to deep learning-based segmentation models, U-Net, FCN, V-Net, Mask-RCNN and YOLO.