Biomarker Localization From Deep Learning Regression Networks.

Biomarker estimation methods from medical images have traditionally followed a segment-and-measure strategy. Deep-learning regression networks have changed such a paradigm, enabling the direct estimation of biomarkers in databases where segmentation masks are not present. While such methods achieve high performance, they operate as a black-box. In this work, we present a novel deep learning network structure that, when trained with only the value of the biomarker, can perform biomarker regression and the generation of an accurate localization mask simultaneously, thus enabling a qualitative assessment of the image locus that relates to the quantitative result. We showcase the proposed method with three different network structures and compare their performance against direct regression networks in four different problems: pectoralis muscle area (PMA), subcutaneous fat area (SFA), liver mass area in single slice computed tomography (CT), and Agatston score estimated from non-contrast thoracic CT images (CAC). Our results show that the proposed method improves the performance with respect to direct biomarker regression methods (correlation coefficient of 0.978, 0.998, and 0.950 for the proposed method in comparison to 0.971, 0.982, and 0.936 for the reference regression methods on PMA, SFA and CAC respectively) while achieving good localization (DICE coefficients of 0.875, 0.914 for PMA and SFA respectively, p < 0.05 for all pairs). We observe the same improvement in regression results comparing the proposed method with those obtained by quantify the outputs using an U-Net segmentation network (0.989 and 0.951 respectively). We, therefore, conclude that it is possible to obtain simultaneously good biomarker regression and localization when training biomarker regression networks using only the biomarker value.

LOCALIZING IMAGE-BASED BIOMARKER REGRESSION WITHOUT TRAINING MASKS: A NEW APPROACH TO BIOMARKER DISCOVERY.

Biomarker inference from biomedical images is one of the main tasks of medical image analysis. Standard techniques follow a segmentation-and-measure strategy, where the structure is first segmented and then the measurement is performed. Recent work has shown that such strategy could be replaced by a direct regression of the biomarker value in using regression networks. While achieving high correlation coefficients, such techniques operate as a 'black-box', not offering quality-control images. We present a methodology to regress the biomarker from the image while simultaneously computing the quality control image. Our proposed methodology does not require segmentation masks for training, but infers the segmentations directly from the pixels that used to compute the biomarker value. The network proposed consists of two steps: a segmentation method to an unknown reference and a summation method for the biomarker estimation. The network is optimized using a dual loss function, L2 for the biomarkers and an L1 to enforce sparsity. We showcase our methodology in the problem of pectoralis muscle area (PMA) and subcutaneous fat area (SFA) inference in a single slice from chest-CT images. We use a database of 7000 cases to which only the value of the biomarker is known for training and a test set of 3000 cases with both, biomarkers and segmentations. We achieve a correlation coefficient of 0.97 for PMA and 0.98 for SFA with respect to the reference standard. The average DICE coefficient is of 0.88 (PMA) and 0.89 (SFA). Comparing with standard segment-and-measure techniques, we achieve the same correlation for the biomarkers but smaller DICE coefficients in segmentation. Such is of little surprise, since segmentation networks are the upper limit of performance achievable, and we are not using segmentation masks for training. We can conclude that it is possible to infer segmentation masks from biomarker regression networks.

Automated Agatston Score Computation in non-ECG Gated CT Scans Using Deep Learning.

INTRODUCTION: The Agatston score is a well-established metric of cardiovascular disease related to clinical outcomes. It is computed from CT scans by a) measuring the volume and intensity of the atherosclerotic plaques and b) aggregating such information in an index. OBJECTIVE: To generate a convolutional neural network that inputs a non-contrast chest CT scan and outputs the Agatston score associated with it directly, without a prior segmentation of Coronary Artery Calcifications (CAC). MATERIALS AND METHODS: We use a database of 5973 non-contrast non-ECG gated chest CT scans where the Agatston score has been manually computed. The heart of each scan is cropped automatically using an object detector. The database is split in 4973 cases for training and 1000 for testing. We train a 3D deep convolutional neural network to regress the Agatston score directly from the extracted hearts. RESULTS: The proposed method yields a Pearson correlation coefficient of r = 0.93; p ≤ 0.0001 against manual reference standard in the 1000 test cases. It further stratifies correctly 72.6% of the cases with respect to standard risk groups. This compares to more complex state-of-the-art methods based on prior segmentations of the CACs, which achieve r = 0.94 in ECG-gated pulmonary CT. CONCLUSIONS: A convolutional neural network can regress the Agatston score from the image of the heart directly, without a prior segmentation of the CACs. This is a new and simpler paradigm in the Agatston score computation that yields similar results to the state-of-the-art literature.

On the Relevance of the Loss Function in the Agatston Score Regression from Non-ECG Gated CT Scans.

In this work, we evaluate the relevance of the choice of loss function in the regression of the Agatston score from 3D heart volumes obtained from non-contrast non-ECG gated chest computed tomography scans. The Agatston score is a well-established metric of cardiovascular disease, where an index of coronary artery disease (CAD) is computed by segmenting the calcifications of the arteries and multiplying each calcification by a factor related to their intensity and their volume, creating a final aggregated index. Recent work has automated such task with deep learning techniques, even skipping the segmentation step and performing a direct regression of the Agatston score. We study the effect of the choice of the loss function in such methodologies. We use a large database of 6983 CT scans to which the Agatston score has been manually computed. The dataset is split into a training set and a validation set of n = 1000. We train a deep learning regression network using such data with different loss functions while keeping the structure of the network and training parameters constant. Pearson correlation coefficient ranges from 0.902 to 0.938 depending on the loss function. Correct risk group assignment measurements range between 59.5% and 81.7%. There is a trade-off between the accuracy of the Pearson correlation coefficient and the risk group measurement, which leads to optimize for one or the other.