Automated medical image quality control system
An automated image quality control system is provided. The system comprises: a first component including a first deep learning network model for estimating a quality of medical images acquired by an imaging device; and a second component for determining whether the medical images are out of a distribution of a training dataset used for training the first deep learning network model.
1 . An automated image quality control system comprising:
(a) a first component including a first deep learning network model trained to output a quality score indicative of a quality level of medical images acquired by an imaging device, wherein the medical images are at least one of the following imaging modalities: Positron Emission Tomography (PET), Magnetic Resonance Imaging (MRI), ultrasound, X-ray imaging, Computed Tomography (CT) and wherein the first deep learning network model comprises a plurality of intermediate layers;
(b) a second component implemented by one or more of the plurality of intermediate layers, wherein the second component is trained to detect one or more of the medical images are out of a distribution of a training dataset used for training the first deep learning network model (out-of-distribution) thereby distinguishing an anomalous caused by out of the distribution from a low quality of the input medical images; and
(c) a controller in communication with the imaging device, and upon receiving the quality score from the first component or a detection of the out-of-distribution from the second component, automatically adjust one or more acquisition parameters for acquiring medical images using the imaging device.
2 . The automated image quality control system of claim 1 , wherein the second component is configured to detect the out-of-distribution using one or more intermediate layer activations of the first deep learning network model during the same forward pass of the first deep learning network model that produces the quality score.
3 . The automated image quality control system of claim 2 , wherein the second component is configured to (i) extract one or more feature maps from the one or more intermediate layers, (ii) for each extracted feature map generate a distance equal to a Mahalanobis distance metric between the feature map and a corresponding reference distribution of feature-map activation previously obtained from the training dataset used to train the first deep learning network model.
4 . The automated image quality control system of claim 3 , wherein the second component is configured to further (iii) compute a combined distance score as a weighted sum of the computed Mahalanobis distance metrics, and (iv) compare the combined distance score with a threshold to determine whether the medical images are out of the distribution of the training dataset.
5 . The automated image quality control system of claim 1 , wherein the training dataset excludes medical images that are out of distribution.
6 . The automated image quality control system of claim 1 , wherein upon detecting one or more of the medical images are out of the distribution of the training dataset, the one or more of the medical images are removed from the medical images to be processed by the first component.
7 . The automated image quality control system of claim 1 , wherein the training dataset includes low quality image data simulated using high quality image data.
8 . The automated image quality control system of claim 4 , further comprising a third component for estimating a registration quality.
9 . The automated image quality control system of claim 8 , wherein estimating the registration quality comprises generating a score indicative of an alignment level between the medical images.
10 . The automated image quality control system of claim 9 , wherein the medical images are acquired using different acquisition parameters.
11 . The automated image quality control system of claim 9 , wherein the score is generated using a second deep learning model.
12 . The automated image quality control system of claim 8 , further comprising a fourth component for calibrating the automated image quality control system to adapt to different tasks, sites, scanners, or quality standards.
13 . The automated image quality control system of claim 12 , wherein calibrating the automated image quality control system comprises receiving a user input indicative of site-specific, scanner-specific, or task-specific quality preferences and, in response to the received user input, adjusting at least one of: a threshold used by the second component to detect the out-of-distribution, a weight used to compute the combined distance score, and a mapping between the quality score and a recommended acquisition parameter.
14 . The automated image quality control system of claim 12 , wherein calibrating the automated image quality control system is performed automatically without user intervention.
15 . The automated image quality control system of claim 12 , wherein the fourth component is configured to distinguish between the different tasks, sites, scanners, or quality standards based at least in part on the registration quality estimated by the third component and selects a corresponding pre-calibrated parameter set for the first deep learning network model from a plurality of stored calibration parameter sets.
16 . The automated image quality control system of claim 1 , wherein the controller comprises a fifth component for generating real-time feedback for adjusting the one or more acquisition parameters.
17 . The automated image quality control system of claim 16 , wherein the real-time feedback comprises an instruction to the imaging device to reacquire a medical image with the one or more adjusted acquisition parameters.
18 . A computer-implemented method for automating image quality control comprising:
(a) providing a first component including a first deep learning network model trained to output a quality score indicative of a quality level of medical images acquired by an imaging device, wherein the medical images are at least one of the following imaging modalities: Positron Emission Tomography (PET), Magnetic Resonance Imaging (MRI), ultrasound, X-ray imaging, Computed Tomography (CT) and wherein the first deep learning network model comprises a plurality of intermediate layers;
(b) providing a second component implemented by one or more of the plurality of intermediate layers, wherein the second component is trained to detect one or more of the medical images are out of a distribution of a training dataset used for training the first deep learning network model (out-of-distribution) thereby distinguishing an anomalous caused by out of the distribution from a low quality of the input medical images; and
(c) upon receiving, by a controller in communication with the imaging device, the quality score from the first component or a detection of the out-of-distribution from the second component, automatically adjusting one or more acquisition parameters for acquiring medical images using the imaging device.
19 . The computer-implemented method of claim 18 , wherein the second component is configured to detect the out-of-distribution using one or more intermediate layer activations of the first deep learning network model during the same forward pass of the first deep learning network model that produces the quality score.