Improving Lesion Segmentation in FDG-18 Whole-Body PET/CT scans using Multilabel approach: AutoPET II challenge
Authors:
Gowtham Krishnan Murugesan,
Diana McCrumb,
Eric Brunner,
Jithendra Kumar,
Rahul Soni,
Vasily Grigorash,
Stephen Moore,
Jeff Van Oss
Abstract:
Automatic segmentation of lesions in FDG-18 Whole Body (WB) PET/CT scans using deep learning models is instrumental for determining treatment response, optimizing dosimetry, and advancing theranostic applications in oncology. However, the presence of organs with elevated radiotracer uptake, such as the liver, spleen, brain, and bladder, often leads to challenges, as these regions are often misiden…
▽ More
Automatic segmentation of lesions in FDG-18 Whole Body (WB) PET/CT scans using deep learning models is instrumental for determining treatment response, optimizing dosimetry, and advancing theranostic applications in oncology. However, the presence of organs with elevated radiotracer uptake, such as the liver, spleen, brain, and bladder, often leads to challenges, as these regions are often misidentified as lesions by deep learning models. To address this issue, we propose a novel approach of segmenting both organs and lesions, aiming to enhance the performance of automatic lesion segmentation methods. In this study, we assessed the effectiveness of our proposed method using the AutoPET II challenge dataset, which comprises 1014 subjects. We evaluated the impact of inclusion of additional labels and data in the segmentation performance of the model. In addition to the expert-annotated lesion labels, we introduced eight additional labels for organs, including the liver, kidneys, urinary bladder, spleen, lung, brain, heart, and stomach. These labels were integrated into the dataset, and a 3D UNET model was trained within the nnUNet framework. Our results demonstrate that our method achieved the top ranking in the held-out test dataset, underscoring the potential of this approach to significantly improve lesion segmentation accuracy in FDG-18 Whole-Body PET/CT scans, ultimately benefiting cancer patients and advancing clinical practice.
△ Less
Submitted 2 November, 2023;
originally announced November 2023.
The AIMI Initiative: AI-Generated Annotations for Imaging Data Commons Collections
Authors:
Gowtham Krishnan Murugesan,
Diana McCrumb,
Mariam Aboian,
Tej Verma,
Rahul Soni,
Fatima Memon,
Keyvan Farahani,
Linmin Pei,
Ulrike Wagner,
Andrey Y. Fedorov,
David Clunie,
Stephen Moore,
Jeff Van Oss
Abstract:
The Image Data Commons (IDC) contains publicly available cancer radiology datasets that could be pertinent to the research and development of advanced imaging tools and algorithms. However, the full extent of its research capabilities is limited by the fact that these datasets have few if any, annotations associated with them. Through this study with the Artificial Intelligence (AI) AI in Medical…
▽ More
The Image Data Commons (IDC) contains publicly available cancer radiology datasets that could be pertinent to the research and development of advanced imaging tools and algorithms. However, the full extent of its research capabilities is limited by the fact that these datasets have few if any, annotations associated with them. Through this study with the Artificial Intelligence (AI) AI in Medical Imaging (AIMI) initiative, we produced high-quality, AI-generated imaging annotations of tissues, organs, and/or cancers for 11 distinct medical imaging collections from the IDC. These collections have a variety of image modalities, computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) imaging modalities. Furthermore, the imaging collections cover various body parities, such as the chest, breast, kidneys, prostate, and liver. Both publicly available and novel AI algorithms were adopted and further developed using open-sourced data coupled with expert annotations to create the AI-generated annotations. A portion of the AI annotations were reviewed and corrected by a radiologist to assess the AI model's performance. Both the AI's and the radiologist's annotations conformed to DICOM standards for seamless integration into the IDC collections as third-party analyses. All the models and datasets (images and annotations) are publicly accessible.
△ Less
Submitted 21 November, 2023; v1 submitted 23 October, 2023;
originally announced October 2023.
