TLDR: A new study introduces an automated deep learning method using nnU-Net to segment hepatic anatomy from MRI scans for surgical planning. This AI-powered approach accurately delineates liver parenchyma, tumors, and vascular structures, significantly reducing the time for creating 3D patient-specific models from hours to minutes. The method demonstrated high accuracy, even detecting tumors initially missed by radiologists, making advanced 3D planning more efficient and accessible for liver surgery.
Preoperative planning for liver surgery is a complex process that greatly benefits from detailed 3D models of a patient’s liver anatomy. These models help surgeons understand the spatial relationships between tumors, blood vessels, and other critical structures, aiding in better decision-making before and during operations. Traditionally, creating these 3D models involves manually outlining anatomical structures from diagnostic scans like MRI, which is a time-consuming and labor-intensive task requiring specialized expertise.
A recent study introduces a deep learning-based method to automate this crucial step, aiming to make preoperative planning more efficient and widely accessible. The research, titled “Automated surgical planning with nnU-Net: delineation of the anatomy in hepatobiliary phase MRI”, was conducted by Karin A. Olthof, Matteo Fusaglia, Bianca Güttner, Tiziano Natali, Bram Westerink, Stefanie Speidel, Theo J.M. Ruers, Koert F.D. Kuhlmann, and Andrey Zhylka. Their work focuses on automating the segmentation of hepatic anatomy, including the liver parenchyma, tumors, portal vein, hepatic vein, and biliary tree, from gadoxetic acid-enhanced MRI scans.
The researchers utilized a deep learning network called nnU-Net (version 1) for this purpose. They trained the network on MRI scans from 72 patients, with a particular emphasis on accurately delineating thin structures and preserving the natural tree-like structure of vessels. The performance of this automated method was then evaluated on a separate set of 18 patients by comparing the AI-generated segmentations with those done manually by experts. Additionally, to assess its real-world clinical utility, 10 automated segmentations were generated and then manually refined for actual surgical planning, allowing the researchers to quantify the adjustments needed.
The results were highly promising. The automated system achieved excellent accuracy for liver parenchyma segmentation, with a Dice similarity coefficient (DSC) of 0.97. For other vital structures, the DSCs were 0.80 for the hepatic vein, 0.79 for the biliary tree, 0.77 for tumors, and 0.74 for the portal vein. The average tumor detection rate was 76.6%, with a median of only one false-positive per patient. When integrated into the clinical workflow, the automated segmentations required only minor manual adjustments, especially for the liver parenchyma, portal vein, and hepatic vein, which showed very high DSCs (1.00, 0.98, and 0.95 respectively) after refinement. Tumor segmentation, while more variable, still achieved a DSC of 0.80 after adjustments.
One of the most significant benefits observed during the prospective clinical use of this model was its ability to detect three additional sub-centimeter tumors that radiologists had initially missed. These lesions were later confirmed as malignant by a board-certified radiologist. Furthermore, the integration of this automated segmentation framework dramatically reduced the time required for creating 3D liver models, cutting down the overall processing time from several hours to approximately 15 minutes per patient.
Also Read:
- Advanced AI for Precise Pathological Liver Segmentation
- Automated Dental Structure Mapping in CBCT Scans for Enhanced Patient Care
This advancement means that detailed 3D planning, which was previously only available in highly specialized centers due to the required expertise and resources, can now be applied efficiently as a standard-of-care for every patient undergoing liver surgery. While challenges remain, particularly in fully automating tumor segmentation due to their variable appearance and the difficulty in distinguishing them from benign lesions, this nnU-Net-based method represents a significant step forward in making precise, patient-specific surgical planning more accessible and efficient. For more details, you can refer to the full research paper.
