Image Level Training and Prediction: Intracranial Hemorrhage Identification in 3D Non-Contrast CT

Abstract

Current hardware restrictions pose limitations on the use of convolutional neural networks for medical image analysis. There is a large trade-off between network architecture and input image size. For this reason, identification and classification tasks are commonly approached with patch or region-based methods often utilizing only local contextual information during training and at inference. Here, a method is presented for the identification of intracranial hemorrhage (ICH) in three-dimensional (3D) non-contrast computed tomography (CT). The method combines a convolutional neural network and recurrent neural network in the form of bidirectional long short-term memory (LSTM) for ICH identification at image level. A convolutional neural network is trained for the identification of ICH in axial slices. LSTM is used to analyze the sequential information obtained from slice level classifications. The method is trained end-to-end using full high-resolution 3D non-contrast CTs. At inference, it produces a binary classification with respect to the presence of ICH. A total of 1554 cranial CTs were used to train and validate the method and a separate dataset of 386 images was used for testing. Quantitative analysis showed an area under receiver operating characteristic curve of 0.96. The average time to classification was approximately 0.5 s. Classification of whole 3D images is therefore possible without the need for pre-processing.

Accurate and Efficient Intracranial Hemorrhage Detection and Subtype Classification in 3D CT Scans with Convolutional and Long Short-Term Memory Neural Networks

“…After the initial success of deep learning [ 10 ] in object recognition from images [ 3 , 11 ], deep neural networks have been adopted for a broad range of tasks in medical imaging, ranging from cell segmentation [ 12 ] and cancer detection [ 13 , 14 , 15 , 16 , 17 ] to intracranial hemorrhage detection [ 5 , 8 , 18 , 19 , 20 , 21 , 22 ] and CT/MRI super-resolution [ 23 , 24 , 25 , 26 ]. Since we address the task of intracranial hemorrhage detection, we consider related works that are focused on the same task as ours [ 5 , 6 , 7 , 8 , 18 , 19 , 20 , 21 , 22 , 27 , 28 , 29 , 30 ], as well as works that study intracranial hemorrhage segmentation [ 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 43 , 44 ].…”

“…While most of the recent works proposed deep learning approaches such as convolutional neural networks [ 18 , 20 , 21 , 22 , 27 , 29 , 30 , 37 ], fully-convolutional networks (FCNs) [ 19 , 32 , 33 , 36 , 38 , 39 ] and hybrid convolutional and recurrent models [ 5 , 6 , 7 , 8 ], there are still some recent works based on conventional machine learning methods, e.g., superpixels [ 43 , 44 ], fuzzy C-means [ 31 , 35 ], level set [ 42 , 43 ], histogram analysis [ 41 ], thresholding [ 40 ] and continuous max-flow [ 34 ].…”

“…Their method performs segmentation at multiple levels of granularity and binary classification of intracranial hemorrhage. Another method based on a similar architecture is proposed by Patel et al [ 6 ]. The authors trained the model on a larger dataset of over 25 thousand slices.…”

“…In this paper, we propose an accurate and efficient system based on a lightweight deep neural network architecture composed of a convolutional neural network (CNN) [ 2 , 3 ] that takes as input individual CT slices, and a Long Short-Term Memory (LSTM) network [ 4 ] that takes as input multiple feature embeddings (corresponding to an entire CT scan) provided by the CNN. To our knowledge, there are only a handful of works that employed a similar neural architecture, based on convolutional and recurrent layers, for intracranial hemorrhage detection [ 5 , 6 , 7 , 8 ]. Among these, there is only one prior work that classifies intracranial hemorrhage into subtypes [ 8 ].…”

Integrating uncertainty in deep neural networks for MRI based stroke analysis

“…While the primary interest of our research was not on the development of neural network architectures but rather on integrating uncertainty in the common analysis procedure, our models achieved state-ofthe art results with accuracies of more than 95% at the image-and the patient-level. In future applications, the CNN and the aggregation models may be combined to diagnose patients in an end-toend fashion as for example done in (Patel, et al, 2019), to improve performance. Furthermore, incorporating more than only the neighboring images for feature learning and patient data when combining the image-predictions to patient diagnoses may be of additional interest.…”