Evaluation of deep learning models for quality control of MR spectra.

Purpose: While 3D MR spectroscopic imaging (MRSI) provides valuable spatial metabolic information, one of the hurdles for clinical translation is its interpretation, with voxel-wise quality control (QC) as an essential and the most time-consuming step. This work evaluates the accuracy of machine learning (ML) models for automated QC filtering of individual spectra from 3D healthy control and patient datasets. Methods: A total of 53 3D MRSI datasets from prior studies (30 neurological diseases, 13 brain tumors, and 10 healthy controls) were included in the study. Three ML models were evaluated: a random forest classifier (RF), a convolutional neural network (CNN), and an inception CNN (ICNN) along with two hybrid models: CNN + RF, ICNN + RF. QC labels used for training were determined manually through consensus of two MRSI experts. Normalized and cropped real-valued spectra was used as input. A cross-validation approach was used to separate datasets into training/validation/testing sets of aggregated voxels. Results: All models achieved a minimum AUC of 0.964 and accuracy of 0.910. In datasets from neurological disease and controls, the CNN model produced the highest AUC (0.982), while the RF model achieved the highest AUC in patients with brain tumors (0.976). Within tumor lesions, which typically exhibit abnormal metabolism, the CNN AUC was 0.973 while that of the RF was 0.969. Data quality inference times were on the order of seconds for an entire 3D dataset, offering drastic time reduction compared to manual labeling. Conclusion: ML methods accurately and rapidly performed automated QC. Results in tumors highlights the applicability to a variety of metabolic conditions.

Intraoperative identification of functional brain areas with RGB imaging using statistical parametric mapping: Simulation and clinical studies.

Complementary technique to preoperative fMRI and electrical brain stimulation (EBS) for glioma resection could improve dramatically the surgical procedure and patient care. Intraoperative RGB optical imaging is a technique for localizing functional areas of the human cerebral cortex that can be used during neurosurgical procedures. However, it still lacks robustness to be used with neurosurgical microscopes as a clinical standard. In particular, a robust quantification of biomarkers of brain functionality is needed to assist neurosurgeons. We propose a methodology to evaluate and optimize intraoperative identification of brain functional areas by RGB imaging. This consist in a numerical 3D brain model based on Monte Carlo simulations to evaluate intraoperative optical setups for identifying functional brain areas. We also adapted fMRI Statistical Parametric Mapping technique to identify functional brain areas in RGB videos acquired for 12 patients. Simulation and experimental results were consistent and showed that the intraoperative identification of functional brain areas is possible with RGB imaging using deoxygenated hemoglobin contrast. Optical functional identifications were consistent with those provided by EBS and preoperative fMRI. We also demonstrated that a halogen lighting may be particularity adapted for functional optical imaging. We showed that an RGB camera combined with a quantitative modeling of brain hemodynamics biomarkers can evaluate in a robust way the functional areas during neurosurgery and serve as a tool of choice to complement EBS and fMRI.

A Weakly Supervised Gradient Attribution Constraint for Interpretable Classification and Anomaly Detection.

The lack of interpretability of deep learning reduces understanding of what happens when a network does not work as expected and hinders its use in critical fields like medicine, which require transparency of decisions. For example, a healthy vs pathological classification model should rely on radiological signs and not on some training dataset biases. Several post-hoc models have been proposed to explain the decision of a trained network. However, they are very seldom used to enforce interpretability during training and none in accordance with the classification. In this paper, we propose a new weakly supervised method for both interpretable healthy vs pathological classification and anomaly detection. A new loss function is added to a standard classification model to constrain each voxel of healthy images to drive the network decision towards the healthy class according to gradient-based attributions. This constraint reveals pathological structures for patient images, allowing their unsupervised segmentation. Moreover, we advocate both theoretically and experimentally, that constrained training with the simple Gradient attribution is similar to constraints with the heavier Expected Gradient, consequently reducing the computational cost. We also propose a combination of attributions during the constrained training making the model robust to the attribution choice at inference. Our proposition was evaluated on two brain pathologies: tumors and multiple sclerosis. This new constraint provides a more relevant classification, with a more pathology-driven decision. For anomaly detection, the proposed method outperforms state-of-the-art especially on difficult multiple sclerosis lesions segmentation task with a 15 points Dice improvement.

MRI-based radiomics to predict lipomatous soft tissue tumors malignancy: a pilot study.

OBJECTIVES: To develop and validate a MRI-based radiomic method to predict malignancies in lipomatous soft tissue tumors. METHODS: This retrospective study searched in the database of our pathology department, data from patients with lipomatous soft tissue tumors, with histology and gadolinium-contrast enhanced T1w MR images, obtained from 56 centers with non-uniform protocols. For each tumor, 87 radiomic features were extracted by two independent observers to evaluate the inter-observer reproducibility. A reduction of learning base dimension was performed from reproducibility and relevancy criteria. A model was subsequently prototyped using a linear support vector machine to predict malignant lesions. RESULTS: Eighty-one subjects with lipomatous soft tissue tumors including 40 lipomas and 41 atypical lipomatous tumors or well-differentiated liposarcomas with fat-suppressed T1w contrast enhanced MR images available were retrospectively enrolled. Based on a Pearson's correlation coefficient threshold at 0.8, 55 out of 87 (63.2%) radiomic features were considered reproducible. Further introduction of relevancy finally selected 35 radiomic features to be integrated in the model. To predict malignant tumors, model diagnostic performances were as follow: AUROC = 0.96; sensitivity = 100%; specificity = 90%; positive predictive value = 90.9%; negative predictive value = 100% and overall accuracy = 95.0%. CONCLUSION: This work demonstrates that radiomics allows to predict malignancy in soft tissue lipomatous tumors with routinely used MR acquisition in clinical oncology. These encouraging results need to be further confirmed in an external validation population.

Repetitive motion compensation for real time intraoperative video processing.

In this paper, we present a motion compensation algorithm dedicated to video processing during neurosurgery. After craniotomy, the brain surface undergoes a repetitive motion due to the cardiac pulsation. This motion as well as potential video camera motion prevent accurate video analysis. We propose a dedicated motion model where the brain deformation is described using a linear basis learned from a few initial frames of the video. As opposed to other works using linear basis for the flow, the camera motion is explicitly accounted in the transformation model. Despite the nonlinear nature of our model, all the motion parameters are robustly estimated all at once, using only one singular value decomposition (SVD), making our procedure computationally efficient. A Lagrangian specification of the flow field ensures the stability of the method. Experiments on in vivo data are presented to evaluate the capacity of the method to cope with occlusion or camera motion. The method we propose satisfies the intraoperative constraints: it is robust to surgical tools occlusions, it works in real time, and it is able to handle large camera viewpoint changes.