Peer-reviewed articles
[1] Automatic Detection and Segmentation of Squamous Cervical Cells in Papanicolaou Smear Images, ICMV 2025 (in press).
This paper introduces a cascade combining convolutional neural networks for candidate cell detection with classical image-processing techniques and k-means segmentation in polar coordinates to obtain controlled segmentation of squamous cells. The method is trained and tested on a new Pap-smear database and achieves a combined success rate above 90% on a test set of 95 slides and >1000 cells from 3 patients, demonstrating the feasibility of robust automated detection-plus-segmentation for suspicious cervical cells.
This paper introduces a cascade combining convolutional neural networks for candidate cell detection with classical image-processing techniques and k-means segmentation in polar coordinates to obtain controlled segmentation of squamous cells. The method is trained and tested on a new Pap-smear database and achieves a combined success rate above 90% on a test set of 95 slides and >1000 cells from 3 patients, demonstrating the feasibility of robust automated detection-plus-segmentation for suspicious cervical cells.
[2] Low-Risk Pap Smear Image Filtering Method via Transfer Learning, ICMV 2025 (in press).
This study presents a versatile testing pipeline and a new conventional (non-LBC) Pap-smear dataset aimed at filtering out patients with low or no malignancy risk, thereby reducing the workload of cytologists. Using transfer learning on a ResNet-101 architecture pre-trained on the Brown Multicellular ThinPrep (BMT) dataset and extended to the proprietary data, the system focuses on high sensitivity to high-risk fields. The resulting models are also intended as educational material for students in domains intersecting with cervical cytology.
This study presents a versatile testing pipeline and a new conventional (non-LBC) Pap-smear dataset aimed at filtering out patients with low or no malignancy risk, thereby reducing the workload of cytologists. Using transfer learning on a ResNet-101 architecture pre-trained on the Brown Multicellular ThinPrep (BMT) dataset and extended to the proprietary data, the system focuses on high sensitivity to high-risk fields. The resulting models are also intended as educational material for students in domains intersecting with cervical cytology.
[3] Automated Cervical Cell Classification on Public and Proprietary Datasets, EHB 2025 (in press).
This article evaluates four CNN architectures (InceptionV3, MobileNetV2, VGG16, DenseNet121) for automatic classification of cervical cells into Normal vs. Suspicious using both the public SIPaKMeD dataset and a new proprietary clinical dataset. A two-stage regime (pre-training + fine-tuning) improves cross-domain generalisation, achieving accuracies of 98–99% on SIPaKMeD and ~86–89% on real clinical data. The work highlights the gap between public and clinical data and confirms the potential of CNNs as decision-support tools in cytology screening.
This article evaluates four CNN architectures (InceptionV3, MobileNetV2, VGG16, DenseNet121) for automatic classification of cervical cells into Normal vs. Suspicious using both the public SIPaKMeD dataset and a new proprietary clinical dataset. A two-stage regime (pre-training + fine-tuning) improves cross-domain generalisation, achieving accuracies of 98–99% on SIPaKMeD and ~86–89% on real clinical data. The work highlights the gap between public and clinical data and confirms the potential of CNNs as decision-support tools in cytology screening.
[4] Automatic Segmentation of Orbital Walls, ICCP 2025. (in press).
This paper proposes a supervised and semi-supervised U-Net-based framework for automatic segmentation of orbital walls in cranial CT scans, achieving Dice ≈0.92 and HD95 ≈2.5 mm. The results demonstrate robust performance under varying imaging conditions and are clinically relevant for trauma reconstruction, reinforcing the team’s expertise in high-precision medical segmentation.
This paper proposes a supervised and semi-supervised U-Net-based framework for automatic segmentation of orbital walls in cranial CT scans, achieving Dice ≈0.92 and HD95 ≈2.5 mm. The results demonstrate robust performance under varying imaging conditions and are clinically relevant for trauma reconstruction, reinforcing the team’s expertise in high-precision medical segmentation.
[5] Automatic Multi-Class Segmentation and Landmark-Based Osteotomy Planning for Hip Arthroplasty, ICCP 2025. (in press)
Here, an automatic pipeline segments 11 femoral sub-structures and detects key anatomical landmarks for osteotomy planning, comparing Attention U-Net and 2D V-Net architectures with Dice >0.96. By coupling multi-class segmentation with landmark detection, the method delivers automatic resection planning, improving accuracy and reproducibility in preoperative planning. This broader work in medical image segmentation and planning is directly leveraged in the design of CerviAssistAI segmentation and risk-scoring modules.
Here, an automatic pipeline segments 11 femoral sub-structures and detects key anatomical landmarks for osteotomy planning, comparing Attention U-Net and 2D V-Net architectures with Dice >0.96. By coupling multi-class segmentation with landmark detection, the method delivers automatic resection planning, improving accuracy and reproducibility in preoperative planning. This broader work in medical image segmentation and planning is directly leveraged in the design of CerviAssistAI segmentation and risk-scoring modules.