Augmenting Neurological Disorder Prediction and Rehabilitation Using Artificial Intelligence

Augmenting Neurological Disorder Prediction and Rehabilitation Using Artificial Intelligence focuses on how the neurosciences can benefit from advances in AI, especially in areas such as medical image analysis for the improved diagnosis of Alzheimer’s disease, early detection of acute neurologic events, prediction of stroke, medical image segmentation for quantitative evaluation of neuroanatomy and vasculature, diagnosis of Alzheimer’s Disease, autism spectrum disorder, and other key neurological disorders. Chapters also focus on how AI can help in predicting stroke recovery, and the use of Machine Learning and AI in personalizing stroke rehabilitation therapy.

Other sections delve into Epilepsy and the use of Machine Learning techniques to detect epileptogenic lesions on MRIs and how to understand neural networks.

• Provides readers with an understanding on the key applications of artificial intelligence and machine learning in the diagnosis and treatment of the most important neurological disorders
• Integrates recent advancements of artificial intelligence and machine learning to the evaluation of large amounts of clinical data for the early detection of disorders such as Alzheimer’s Disease, autism spectrum disorder, Multiple Sclerosis, headache disorder, Epilepsy, and stroke
• Provides readers with illustrative examples of how artificial intelligence can be applied to outcome prediction, neurorehabilitation and clinical exams, including a wide range of case studies in predicting and classifying neurological disorders

• Language: English
• Publisher: Academic Press
• Release date: Feb 23, 2022
• ISBN: 9780323886260

Book preview

1

Intracranial hemorrhage detection and classification using deep learning

Naveen Ojha¹ and Sugata Banerji²,    ¹Machine Intelligence Unit, Indian Statistical Institute, Kolkata, India,    ²Department of Mathematics and Computer Science, Lake Forest College, Lake Forest, IL, United States

Abstract

This chapter deals with the detection and classification of intracranial hemorrhage (ICH), which is defined as bleeding inside the skull. ICH may lead to a hemorrhagic stroke, and while this is a less common type of stroke, it still accounts for 10%–15% of cases. According to the Centers for Disease Control and Prevention, stroke accounted for one in every six people dying from cardiovascular diseases in 2018 and this makes ICH a very serious medical condition. It requires urgent, often surgical treatment, and the chances of the patient’s survival depend heavily on the speed of diagnosis. Traditionally, trained human experts locate and identify the type of ICH by inspecting radiological images, such as computerized tomography (CT) scan or magnetic resonance imaging scan images of the patient’s skull. Although this is a standard procedure in cases of head trauma, or for a patient experiencing acute neurological symptoms, the process is complex and slow. Efforts have been made to automate this process using machine learning algorithms to aid in diagnosis, with varying success. Here we describe the different types of ICH and review the different techniques that have been used to address the problem of detecting and classifying them. We also discuss some of the characteristics of this problem that make it especially challenging. Finally, we describe several approaches that we tried on a rich CT image dataset with the results obtained.

Keywords

Intracranial hemorrhage; CT scans; computer vision; image analysis; deep learning; convolutional neural networks

1.1 Introduction

Intracranial hemorrhage (ICH), defined as bleeding inside the skull, is a serious but relatively common health problem. It accounts for approximately 10%–15% of strokes in the US (Rymer, 2011), where stroke accounts for one in every six people dying from cardiovascular diseases (Centers for Disease Control and Prevention) and is the number five cause of death (American Stroke Association). ICH may be caused by a variety of conditions ranging from trauma, stroke, aneurysm, vascular malformations, high blood pressure, illicit drugs, and blood clotting disorders. The consequences also vary vastly, and depending upon the size, type, and location of the hemorrhage, it may cause anything from a headache to death. Therefore identifying the location and type of any hemorrhage present is a critical step in treating the patient.

ICH is diagnosed through history, physical examination, and, most commonly, noncontrast X-ray computerized tomography (CT) examination of the brain, which discloses the anatomic bleeding location. The role of the radiologist is to detect the hemorrhage, characterize the hemorrhage subtype, its size, and to determine the type of treatment it needs, which could potentially include immediate surgery. While all acute (i.e., new) hemorrhages appear dense (i.e., white) on CT scan images, radiologists primarily determine their subtype based on their location, shape and proximity to other structures. The subtypes of ICH are described in Section 1.2. While typically, a larger hemorrhage site is more serious than a small one, even a small hemorrhage can lead to death since it could indicate another type of serious abnormality, such as cerebral aneurysm.

In recent years, digitization of medical images has brought a paradigm shift in the way we use them. Now doctors can not only record, archive, and retrieve images with relative ease, but can also take help of new computational tools for faster and more accurate decision-making. This is vital in all diseases, but particularly crucial in ICH where fast and accurate diagnosis can save lives. With globally available big digital image datasets in healthcare, unprecedented improvement in computing power, efficient low-cost algorithms, and easy access to cloud-based computing solutions, the medical community is showing growing interest in artificial intelligence. Recently machine learning has shown the ability to perform automated extraction of information from images that is typically not visible to human observers, thereby enhancing the diagnostic and prognostic benefits of patients. Machine-based diagnostics also minimizes errors due to human bias, and enhances accessibility across geographically distant locations.

The rest of this chapter is organized as follows. In the next section, we describe the types of intracranial hemorrhage, followed by a brief literature survey of the field. Next, we describe some challenges that are characteristic of analyzing CT scan images in general, and detecting intracranial hemorrhage in particular, along with some solutions. Then, we describe our attempts at tackling this problem using some established deep learning models. Finally, in the Conclusions section, we summarize our findings.

1.2 Types of intracranial hemorrhage

Fig. 1.1, sourced from Wikipedia (Wikipedia, the free encyclopedia, 2021) and created by user Mysid, based on work by SEER Development Team (Meninges—SEER training, 2021), shows the cross-section of the meninges, or the membranes enclosing the brain in the human cranium. ICH can be classified into Intraparenchymal hemorrhage (IPH), Intraventricular hemorrhage (IVH), Subarachnoid hemorrhage (SAH), Subdural hematoma (SDH) or Epidural hematoma (EDH) based on the location of bleeding with respect to these membranes, bone and the brain matter. We briefly describe these types below. It should be noted that a patient often suffers from one or more types of ICH at the same time and hence any classification algorithm diagnosing ICH must be able to assign multiple class labels to a CT image.

Figure 1.1 Meninges of the central nervous parts. Figure sourced from Wikipedia and created by user Mysid, based on work by SEER Development Team.

1.2.1 Intraparenchymal hemorrhage

IPH is a form of ICH where the bleeding occurs within brain parenchyma. Although it accounts for only 6.5%–19.6% of all strokes, IPH is associated with the maximum mortality rate, with a one-year survival rate of 40% and a 10-year survival rate of just 24% (Gross, Jankowitz, & Friedlander, 2019). In this type of hemorrhage, the bleeding is located completely within the brain itself. Since it is very likely to result in death or major disability, it constitutes an immediate medical emergency. It can be caused by a variety of causes, but occurs most commonly due to either hypertension or cerebral amyloid angiopathy (Gross et al., 2019). Most accurate diagnosis is done through CT or magnetic resonance imaging (MRI) scans of the brain. An example of a CT scan with IPH is shown in Fig. 1.2A.

Figure 1.2 CT scan images of the brain showing (A) intraparenchymal hemorrhage (IPH), (B) intraventricular hemorrhage (IVH), (C) subarachnoid hemorrhage (SAH), (D) subdural hematoma (SDH) and (E) Epidural hematoma (EDH). The bleeding can be seen as a whitish patch inside the skull (marked by arrows added by us). The types have been described in the text. The images have been sourced from Wikipedia (Wikipedia, the free encyclopedia, 2021). In particular, image (B) is from ( Yadav et al., 2007), (C), (D) and (E) are by James Heilman, MD.

1.2.2 Intraventricular hemorrhage

IVH, like IPH, is another form of ICH where the blood is within the brain parenchyma. Unlike IPH, however, in IVH the bleeding occurs into the brain’s ventricular system where the cerebrospinal fluid is produced and circulates through towards the subarachnoid space. IVH is a particularly serious condition with an expected mortality rate between 50% and 80% (Hinson, Hanley, & Ziai, 2010). Approximately 70% of IVHs are secondary which occur as an extension of an IPH or SAH into the ventricular system (Hinson et al., 2010). Fig. 1.2B shows an example of a CT scan with IVH.

1.2.3 Subarachnoid hemorrhage

SAH is bleeding into the subarachnoid space, which is the area between the arachnoid membrane and the pia mater surrounding the brain. It accounts for only 5% of strokes but can occur at a relatively young age. Symptoms of SAH may include sudden and severe headaches, vomiting, fever, decreased level of consciousness, and occasionally seizures (Abraham & Chang, 2016). SAH may be caused by head injury or spontaneously, usually from a ruptured cerebral aneurysm. Diagnosis is done by CT imaging, and sometimes, a lumbar puncture is also needed (Carpenter, Hussain, & Ward, 2016). An example of a CT scan with SAH is shown in Fig. 1.2C.

1.2.4 Subdural hematoma

An SDH occurs when a blood vessel bursts near the surface of the brain and blood accumulates between the arachnoid mater and the dura mater, which are the two outer membranes surrounding the brain. It is most commonly caused by a tear in a blood vessel vein in the subdural space as a result of a traumatic brain injury (TBI). The symptoms are varied and may include confused or slurred speech, problems with balance, headache, nausea, and loss of consciousness (Kotwica & Brzeziński, 1993). Emergency surgery may be needed to treat an SDH. An example of a CT scan with SDH can be seen in Fig. 1.2D.

1.2.5 Epidural hematoma

An EDH is bleeding that occurs between the dura mater and the skull. It is typically caused by a burst artery due to a TBI. EDH often causes a loss of consciousness, headaches, confusion, vomiting, an inability to move parts of the body, and occasionally seizures (Ferri, 2019). They occur in about 10% of TBIs and diagnosis is typically done using a CT scan or an MRI, the latter being more accurate. It is a surgical emergency and usually results in death if surgery is not performed (Khairat & Waseem, 2021). Fig. 1.2E displays an example of a CT scan with EDH.

1.3 Related work

Deep learning is a kind of representation learning, where a computer program learns from raw training data to automatically discover the representations needed for detection or classification (Lecun, Bengio, & Hinton, 2015). It uses multilayered neural networks, and very complex functions can be learned by the combination of enough such layers. In the context of medical images, deep learning uses the raw pixel values of the images rather than extracted features as input. This helps the system to avoid errors caused by inaccurate segmentation and/or subsequent feature extraction.

Convolutional neural networks (CNNs or ConvNets) constitute one of the popular models of deep learning. The breakthrough in CNNs came with the ImageNet competition in 2012 (Krizhevsky, Sutskever, & Hinton, 2012), where the error rate was almost halved for object recognition. CNNs were revolutionally revived through the efficient use of Graphics Processing Units (GPUs) (Simonyan & Zisserman, 2014), Rectified Linear Units (ReLUs) (Glorot, Bordes, & Bengio, 2011), dropout regularization (Srivastava, Hinton, Krizhevsky, Sutskever, & Salakhutdinov, 2014) and data augmentation (Krizhevsky et al., 2012). CNNs come in all shapes and sizes but they usually have some elements in common. The structure of MobileNetV2 (Sandler, Howard, Zhu, Zhmoginov, & Chen, 2018), one of the classification networks used in this work, is shown in Fig. 1.3. This is just one of many possible architectures used for classification, segmentation, feature extraction, and a variety of other specialized tasks.

Figure 1.3 Structure of a convolutional neural networks (CNN). The structure of MobileNetV2, one of the convolutional neural networks used in this work. The representative input image shown has been sourced from Wikipedia (Wikipedia, the free encyclopedia, 2021)

Advances in deep learning have shown great promise towards extracting clinically important information from medical images. Some examples include detection of metastases in histologic sections of lymph nodes (Ehteshami Bejnordi, Veta, & Johannes van Diest, 2017), grading of diabetic retinopathy on retinal fundus photographs (Gulshan, Peng, & Coram, 2016), diagnosis of thoracic ailments from chest radiographs (Rajpurkar, Irvin, & Ball, 2018), and classification of images of skin cancer (Esteva, Kuprel, & Novoa, 2017), with accuracies comparable to or, in some cases, exceeding that of experts. Great progress has been made in digital histopathology as well, with deep learning algorithms being used for automated annotation and classification of cancer from tissue slide images (Komura & Ishikawa, 2018).

Recently, deep learning has been used to analyze brain CT images with great success (Gao, Hui, & Tian, 2017; Grewal, Srivastava, Kumar, & Varadarajan, 2018). A patch-based CNN has achieved an ICH detection performance exceeding that of 2 of 4 American Board of Radiology-certified radiologists (Kuo, Hӓne, Mukherjee, Malik, & Yuh, 2019). However, some of the inherent limitations of deep learning include high computational cost and requirement of a large number of accurately labeled training images (Litjens, Kooi, & Bejnordi, 2017). A large and diversely sourced training dataset can vastly improve the generalization performance of deep learning algorithms (Sun, Shrivastava, Singh, & Gupta, 2017). With an increase in the number of well-annotated training images available in the public domain in the coming days, we can expect to see even better results from deep learning algorithms in brain CT image analysis in general, and ICH diagnosis in particular.

1.4 Characteristic challenges of intracranial hemorrhage detection from computerized tomography images

In this section, we first briefly describe the CT imaging process and the file format used to store the images. Then, we discuss some of the challenges that are faced while analyzing these images.

1.4.1 The computerized tomography imaging process

CT scans (formerly known as computed axial tomography or CAT scans) are medical images created by recording multiple X-ray images of a part of the body using a rotating X-ray machine, and then combining these images on a computer. A CT scan image shows a 2-D slice of the body part in question, although several such images could be combined to create a 3-D model. Fig. 1.4 shows a collection of 34 such slices of the human brain, from base of the skull to the top. Since the exposure to X-ray radiation comes with some potential health risks, studies using CT scans cannot generally recruit healthy volunteers or large nonclinical populations, and have to rely on archived data from hospitals, clinical trials and old medical records (Muschelli, 2019). Most of the CT scan data is stored in the form of Digital Imaging and Communications in Medicine (DICOM) files. These files contain a large amount of metadata in addition to the pixel values of the actual image. The image itself is most typically 512×512 pixels in size. It is a single plane image, but has a much higher dynamic range than a regular grayscale image, as discussed later.

Figure 1.4 Slices of the brain. Computerized tomography of human brain, from base of the skull to top. Taken with intravenous contrast medium. Sourced from Wikipedia and uploaded by Mikael Häggström, Radiology, Uppsala University Hospital.

1.4.2 Availability of data

The presence of protected health information (PHI) is a common issue with medical datasets in general, and DICOM files in particular, since the format contains a large amount of metadata in the header where such data could be present. This means a large part of the data is not publicly available for research without proper anonymization (discussed in the next section). Also, because of the lack of scans on healthy subjects for studies as mentioned above, most of the CT data is acquired clinically and not in a research setting (Muschelli, 2019). Where such data is available, such as on Radiopaedia (Radiopaedia, 2021), the images are often converted from DICOM to standard JPEGS and hence lose information on Hounsfield units (discussed below). Among the large DICOM datasets that are publicly available are the CQ500 (Chilamkurthy, Ghosh, & Tanamala, 2018), the Stroke Imaging Repository Consortium (The Stroke Imaging Repository Consortium, 2021), and the Cancer Imaging Archive (TCIA) (The Cancer Imaging Archive, 2021). However, it should be noted that not all the head CT images present in these datasets show cases of ICH. The TCIA dataset, for instance, mostly shows brain cancer cases.

For the experiments described below, we use a large dataset published by the Radiological Society of North America (RSNA) for an ICH detection competition held on Kaggle in 2019 (Radiological Society of North America RSNA Intracranial Hemorrhage Detection, 2021). We discuss this dataset in more detail there.

1.4.3 Digital imaging and communications in medicine anonymization

DICOM files have a header where a large amount of metadata can be embedded. Each DICOM file has a standard set of fields in the header where metadata related to the patient or technical data about the scanned image could be stored. This technical data should ideally be consistent across studies, but many manufacturers and sites do not conform to the standard and store data in nonstandard fields (Muschelli, 2019). The aim of anonymization, therefore is to remove the data in these fields if they contain PHI, but retain the data if it contains relevant technical information of the scan for analysis. Data that does not conform to the DICOM standard presents an extra challenge to this process. This is especially problematic since the PHI data must be read to be deleted, and often, the anonymization needs to be done before the data is read. The RSNA Medical Imaging Resource Community (MIRC) Clinical Trials Processor (Radiological Society of North America RSNA Medical Imaging Resource Community, 2021) and the DICOM library upload service (DICOM Library, 2021) are two DICOM anonymization solutions recommended from several such options reviewed by Aryanto, Oudkerk, and van Ooijen (2015). Muschelli (2019) provides an in-depth discussion about the actual process of reading the DICOM data and converting it to other formats. This topic is out of scope of this current chapter.

1.4.4 Hounsfield values and windowing

CT images are single-plane images like grayscale images, but unlike regular grayscale images, the range of values that the pixels can have is much greater. The individual pixels values in a CT image follow the Hounsfield unit (HU) scale where the radiodensity of distilled water at standard pressure and temperature (STP) is defined as zero HU, while the radiodensity of air at STP is defined as –1000 HU (Tawfik, Abdelhalim, & Elkafrawy, 2012). Other body tissues have varying densities (Fosbinder & Orth, 2012). For diagnosing CT images, the most important tissue types are bone (+300–+1900 HU), brain matter (+20–+46 HU) and blood (+13–+100 HU) (Fosbinder & Orth, 2012; Rao, Singh, Khandelwal, & Sharma, 2016).

While actual HU values present in a CT image may be several thousands, human eyes, and computer displays are not capable of distinguishing so many different shades of gray. Therefore only a limited number of HU can be displayed at any given time. This process is known as windowing. A window level (WL) and window width (WW) suitable for the tissue being studied is chosen. The WW is the number of HU values displayed and the WL is the middle value of this range. Tissues within this range are displayed as various shades of gray. Tissues with HU values outside this range are displayed as either black or white (Tawfik et al., 2012). Fig. 1.5 shows the same CT image with different WL and WW values. Windowing is a very important image preprocessing step and often the first step in analyzing CT images.

Figure 1.5 Windowing. A CT scan showing the effect of different window levels and window widths on the display. (A) has a window level (WL) of 25 and window width (WW) of 95, which shows most of the tissue in the head, but makes the bone and the blood (present outside the skull) very similar in appearance. (B) has a WL of 300 and WW 2500 and shows the details within the bone. (C) has a WL and WW of 60 and 80 respectively, and separates the details of the blood from the both the bone and the brain matter. The brain CT image is downloaded from Medimodel ( Medimodel).

1.5 Our approach

We studied several deep learning techniques for detection and classification of ICH from CT scans and tried several of these techniques on the RSNA Kaggle challenge dataset described below. Each of these techniques is briefly described below. We used the tool MicroDicom (MicroDicom - Free DICOM Viewer & Software, 2021) to preview DICOM files and generate the JPEGs used in the figures in this chapter. We ran our Python code on a laptop with an NVIDIA GPU. It should be noted that we did not develop the models that we use—we merely test them on this problem.

1.5.1 Dataset

The data we use for our experiments and for most of the figures in this chapter is from a very large dataset published by the RSNA. This dataset was released on Kaggle in 2019 for an ICH detection competition (Radiological Society of North America RSNA Intracranial Hemorrhage Detection, 2021). This dataset was compiled by assembling large volumes of anonymized CT data provided by four research institutions, namely, Stanford University, Thomas Jefferson University, Unity Health Toronto and Universidade Federal de São Paulo (UNIFESP). The American Society of Neuroradiology (ASNR) (American Society of Neuroradiology ASNR, 2021) used more than 60 volunteers to label over 25,000 exams for the challenge dataset. Tooling and support for the data annotation process was provided by MD.ai. The platform for medical AI (2021).

The challenge dataset consists of 877, 180 brain CT scan images in the DICOM format, out of which 755, 948 form the training set, and 121, 232 form the test set. Each image is a single 2-D slice of the brain 512×512 pixels in size. Each of these images shows zero or more types of ICH. There is great deal of imbalance in the data, with nearly 86% of the data coming from individuals without any ICH. Among the five types, SDH is seen in the most number of cases (approximately 32%) and EDH occurs in the least number of cases (just 2%). Since each image can contain more than one type of ICH, these images form approximately 4.5 million data points for training a deep learning model.

1.5.2 Experiments

1.5.2.1 Data preprocessing

The first step in the analysis of our CT images was preprocessing. This consisted of removal of bad images which are scans without any part of the brain visible or just a small amount of bone at the top of the skull visible. In the remaining images we did cropping and padding to remove most of the surrounding black region using standard image processing functions, and also to bring the cropped images to the same size. We also tried skull and face removal as suggested by (Kuo et al., 2019). However, several kinds of anomalies in the images created difficulties for this step. For instance, some subjects had part of the skull missing, others had bony structures and scanning artefacts inside the brain matter, or bleeding and swelling outside the skull. Also, the orientation of the head varied slightly between images. These were hard to handle consistently using programs and consequently we got better results without the skull and face removal step.

1.5.2.2 Windowing

We extracted three different windows for blood, brain matter, and bone, and combined them into a three-plane false-color image that contained three times the information as a single-plane image. These false-color images were used for training the neural networks and were found to generate better results. An example of this operation.

1.5.2.3 Multilabel classification

Multilabel classification deals with multiple labels being assigned to every instance in a dataset. That is, an instance can be simultaneously assigned more than one class label. The main baseline of multilabel classification is an effective and computationally efficient technique called Binary relevance, which is a decomposition method based on the learning assumption that labels are independent. Therefore each label is classified as relevant or irrelevant by a binary classifier learned for that label independently from the rest of the labels (Tawiah & Sheng). In our case, we had to train a model that takes a brain CT scan image as input and outputs a vector y whose individual components represent the probabilities of the subject having each of the ICH subtypes including the label any which indicates whether any of the types exist or not.

The loss function we used for this model was binary crossentropy as this has been found to work well in multilabel classification problems with imbalanced datasets (Ho & Wookey, 2020; Rezaei-Dastjerdehei, Mijani, & Fatemizadeh, 2020). Rectified Adam (RAdam), a novel variant of Adam, was used as the optimizing function by introducing a term to rectify the variance the adaptive learning rate (Liu, Jiang & he). RAdam brings consistent improvement over the vanilla Adam (Kingma & Ba). If each data point has exactly one output label, then a softmax activation function is used over the output nodes to make the probabilities add up to 1. But when each data point has multiple output labels, as in our problem, then there isn’t a need for the probability to sum to 1 and we use a sigmoid as the output activation function in the last layer.

1.5.2.4 Network architectures used

We used three network architectures for ICH classification on our data. Because we had a large amount of training data, which we further increased by augmentation, we trained these networks from scratch. We used the same preprocessing methods and other parameters described in the previous sections with all three network architectures.

The first such network that we used was the Resnet50 (He, Zhang, Ren, & Sun, 2016). The motivation for this choice was the perception that the deeper the better when it comes to CNNs. However, one of the bottlenecks of earlier networks was that the performance degraded after a certain depth. ResNet50 uses the residual learning framework to ease the training of networks that are substantially deeper than those used previously, and hence this model was chosen.

The second network architecture that we chose was a new mobile architecture, MobileNetV2 (Sandler et al., 2018). This model has been seen to improve the state of the art performance of mobile models on multiple tasks and benchmarks as well as across a spectrum of different model sizes and this was our motivation for choosing this architecture. Fig. 1.3 shows the structure of this network.

The third and final network that we tested was the EfficientNetB2 (Tan & Le, 2019). Scaled-up ConvNets is widely used to achieve better accuracy. However, the process of scaling up ConvNets has never been well understood and there are currently many ways to do it. The most common way is to scale up ConvNets by their depth or width. Another less common, but increasingly popular, method is to scale up models by image resolution. Though it is possible to scale one or more of these three dimensions arbitrarily, this scaling requires tedious manual tuning and still often yields suboptimal accuracy and efficiency. In EfficientNets (Tan & Le, 2019), the authors study and rethink the process of scaling up ConvNets to develop EfficientNets. EfficientNets are a family of image classification models, which achieve state-of-the-art accuracy, yet being an order-of-magnitude smaller and faster than previous models. Since our computing power was limited, we were motivated to test this model.

1.5.3 Results

The RSNA Kaggle ICH Detection dataset (Radiological Society of North America RSNA Intracranial Hemorrhage Detection, 2021) does not have labels for the test data. Therefore we were unable to know the accuracy of our model at the query image level. We could only test the relative accuracy of the three models that we tested by comparing the scores obtained from the Kaggle site on submitting our results. These results are in the form of a list of image filenames, where each filename is repeated six times (once for each of the five ICH subtypes, and once for any). Each instance of the filename is followed by one probability score for that image having a particular class label. On submission, we get a multilabel logarithmic loss value and a rank among all the teams that participated in the competition when it was held in 2019. These results are tabulated in Table 1.1. It can be seen both from the loss and the rank that the EfficientNetB2 architecture performs the best out of the three, followed by MobileNetV2 and ResNet50.

Table 1.1

1.6 Conclusion

In this chapter, we introduced the problem if ICH detection and classification using deep learning. We started by briefly describing the different types of ICH, then did a brief literature survey and went on to describe the characteristics and challenges of analyzing CT scan images to diagnose ICH. Finally, we combined some preprocessing techniques with three different well-known CNN architectures and applied on a large CT image dataset to detect and classify occurrences of hemorrhage.

While deep learning has been able to diagnose and localize ICH in CT scans with great success, there are clearly areas where there is ample scope for improvement. Preprocessing of images is one such area. While the HU values for different kinds of tissues are known, there is considerable overlap between these values and hence the choice of window is a crucial factor in success. Also, traditional image processing operations are usually used to crop and remove unnecessary parts of the scanned image. This is another area where machine learning in general, and deep learning in particular could bring more success. Finally, the availability of multiple CT slices of the same brain can allow the exploration of other architectures where the 3-D structure of the ICH can be used for diagnosis. Such models have been mostly kept out of the current chapter due to the single-slice nature of the dataset