The quality and richness of feature maps extracted by convolution neural networks (CNNs) and vision Transformers (ViTs) directly relate to the robust model performance. In medical computer vision, these information-rich features are crucial for detecting rare cases within large datasets. This work presents the "Scopeformer," a novel multi-CNN-ViT model for intracranial hemorrhage classification in computed tomography (CT) images. The Scopeformer architecture is scalable and modular, which allows utilizing various CNN architectures as the backbone with diversified output features and pre-training strategies. We propose effective feature projection methods to reduce redundancies among CNN-generated features and to control the input size of ViTs. Extensive experiments with various Scopeformer models show that the model performance is proportional to the number of convolutional blocks employed in the feature extractor. Using multiple strategies, including diversifying the pre-training paradigms for CNNs, different pre-training datasets, and style transfer techniques, we demonstrate an overall improvement in the model performance at various computational budgets. Later, we propose smaller compute-efficient Scopeformer versions with three different types of input and output ViT configurations. Efficient Scopeformers use four different pre-trained CNN architectures as feature extractors to increase feature richness. Our best Efficient Scopeformer model achieved an accuracy of 96.94% and a weighted logarithmic loss of 0.083 with an eight times reduction in the number of trainable parameters compared to the base Scopeformer. Another version of the Efficient Scopeformer model further reduced the parameter space by almost 17 times with negligible performance reduction. In summary, our work showed that the hybrid architectures consisting of CNNs and ViTs might provide the desired feature richness for developing accurate medical computer vision models.
Low-grade gliomas (LGG) invade the brains of young and productive adults leading to significant neurological morbid-ity. Mutations in the isocitrate dehydrogenase (IDH) gene are very common in LGG. Because IDH1 inhibitors are ef-fective in the treatment of LGG, early detection of growth is critical to initiating therapy without delay. This retrospec-tive observational study evaluates the ability of AI-assisted volumetric analysis to correctly detect tumor growth in lon-gitudinal studies of LGG as compared to the standard clinical method of visual inspection by radiologists. The study includes 63 participants categorized into clinical progression (n = 34), clinically stable (n=22), and negative control groups (n = 7). The dates of detection of tumor growth by visual inspection were gathered from radiological reports. Longitudinal tumor volumes were calculated from automated segmentations by the MRIMath FLAIR AI. Golden truths were obtained by physician review using the MRIMath Smart manual contouring system. Growth by significant shifts in tumor volumes was detected by using the statistical method of online change-of-point method. Our results demonstrate that in the clinical progression group, automatic AI segmentation followed by human review detected tu-mor growth at a median of 21 months earlier than visual inspection. In the clinically stable group, AI with human re-view identified growth in 13/22 cases at a median of 23 months earlier than the last MRI. AI without human review generated similar results but with a 25% false positive and an 8.33% false negative rate. The median time spent by physicians in reviewing, revising, and approving the AI segmentations is 2 minutes. These findings highlight the clin-ical potential of AI-assisted volumetric analysis followed by physician oversight for the timely detection of tumor pro-gression in LGG patients.
Glioblastomas (GBM) grow rapidly and infiltrate the cerebral parenchyma, leading to significant neurological morbidity and mortality. Early detection of tumor growth promotes early therapeutic interventions that delay neurological progression and possibly prolong survival times. This retrospective observational study evaluates the ability of AI-assisted volumetric analysis to correctly detect tumor progression in longitudinal studies of newly diagnosed GBM, compared to the standard clinical method of visual inspection by radiologists and neuro-oncologists. Fifteen of 56 patients met the inclusion and exclusion criteria. The dates of tumor progression were gathered from clinical reports. Longitudinal tumor volumes were calculated from automated segmentations by the MRIMath T1c AI followed by physician review using the MRIMath Smart manual contouring system. Growth by significant shifts in tumor volumes was detected by using the statistical method of the online change-of-point method. Our results demonstrate that automatic AI segmentation followed by human review detects tumor progression earlier than clinical notes in 4/15 patients at a median of 105 days. Furthermore, the longitudinal AI-measured volumes validated two cases of pseudoprogression as evidenced by subsequent volume stability. This study emphasizes the enhanced diagnostic accuracy achieved by incorporating volumetric data analysis into clinical decision-making. This can be accomplished by integrating efficient AI-powered segmentation and a streamlined human review system into the clinical workflow.