One of the top 5 disorders that cause death is intracranial hemorrhage. To relieve busy radiologists and identify patients in need of rapid treatment, we attempt to automate the process of identifying these bleeding in this research. we experimented with three different pretrained models, namely ResNet101, DenseNet121, and AlexNet, to improve the accuracy of their classification model. We observed that switching from ResNet50 to ResNet101 led to a modest increase in accuracy from 91.3% to 91.7%. However, using DenseNet121, which has a unique architecture where each layer receives collective knowledge from all previous layers, resulted in the highest accuracy of 91.8%. On the other hand, AlexNet, a well-known architecture that won the 2012 ImageNet competition, had a shorter training period but achieved a lower accuracy of only 89%.
One of the top five global causes of mortality is intracranial hemorrhage, sometimes known as brain hemorrhage. The illness is brought on by bleeding inside the cranium, commonly known as the skull. It is critical to get a prompt and correct diagnosis because this form of hemorrhage accounts for about 10% of strokes in the United States. Effective management of the illness and avoidance of its serious effects depend on an accurate diagnosis, underscoring its crucial importance in patient care.
The magnitude, kind, and location of an intracranial hemorrhage, often known as brain bleeding, can have a variety of effects. These variables interact intricately, and a tiny hemorrhage in a crucial area may be just as lethal as a bigger one. Rapid diagnosis is essential in these situations, especially for those displaying severe hemorrhage signs like unconsciousness. The diagnostic and treatment procedure, however, is frequently challenging and time-consuming, resulting in considerable delays and unfavorable patient outcomes, particularly for those who require urgent care. This study attempts to overcome this difficulty by automating the first classification of the bleed's size, location, and kind using an algorithm. With the algorithm treating every scan similarly to reduce human error and increase accuracy, clinicians are able to diagnose patients more quickly and accurately.
In order to improve the diagnosis process and eventually result in improved patient outcomes, this research intends to create an algorithm capable of properly recognizing acute cerebral hemorrhage and its subtypes. A major step towards overcoming the difficulties involved in diagnosing and treating cerebral hemorrhage will be reached with the effective identification of these essential characteristics. The algorithm will shorten the diagnostic process and lower the chance of mistakes by automating this process, allowing medical practitioners to diagnose patients more quickly and accurately.
II. BACKGROUND AND RELATED WORK
Deep learning has been used more and more in medical imaging during the past ten years, making it possible to identify pneumonia using chest X-rays or retinal to diagnose diabetes. Image segmentation has been used to automate measurements of organs, cell counts, and simulations. It has also been used to color-code images and extract boundary information. In a recent study, the Qure.ai team used convolutional neural networks (CNNs) and natural language processing (NLP) methods on related medical records to identify MRIs using deep learning. The research, which had trouble with 3D scans, examined 313,318 anonymous CT scans from several centers across India. The dataset had a mean age of 43.4 years, and 42.87% of the people were female. The study's overall AUC for all sub-categories was 0.94 ± 0.3. Other important studies include acute cerebral hemorrhage on head computed tomography expert-level identification, extracting 2D weak labels from volume labels using multiple instance learning, and extracting 2D weak labels from volume labels. In a subsequent study, they aggregated the losses per slice derived from bags of 2D slices to update the model's parameters.
A. Data Cleaning
Both picture data and metadata are included in the 194082 DICOM pictures in the collection, with the metadata lacking any private information to protect data privacy. Because using the DICOM format may result in lengthy processing times, the metadata is converted into a data frame. This allows for speedier and more efficient processing. We next show the cross-sectional brain scans so that the dataset may be understood better. Filtering out blank slices is done using the "img pct window" column, which displays the percentage of brain pixels in each slice. Pictures having a value in this column of less than 20% are removed.
 Detecting Intracranial Hemorrhage with Deep Learning Arjun Majumdar et al. Annu Int Conf IEEE Eng Med Biol Soc. 2018 July.
 Samuel W. Remedios, Zihao Wu, Camilo Bermudez, Cailey I. Kerley, Snehashis Roy, Mayur B. Patel, John A. Butman, Bennett A. Landman, and Dzung L. Pham, “Extracting 2D weak labels from volume labels using multiple instance learning in CT hemorrhage detection”
 T.D. Phong, H.N. Duong, H.T. Nguyen, Tran van Hoa, Brain hemorrhage diagnosis by using deep learning, Proceedings of the 2017 International Conference on Machine vii Learning and Soft Computing, 2017.
 Leslie N. Smith, “Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates”.
 Sasank Chilamkurthy, Rohit Ghosh, Swetha Tanamala, Pooja Rao and Qure.ai team. Development and Validation of Deep Learning Algorithms for Detection of Critical Findings in Head CT Scans.
 Hyunkwang Lee Sehyo, Yune Mohammad, Mansouri Myeongchan Kim, Shahein H. Tajmir, Claude E. Guerrier, A Ebert Sarah, R Pomerantz Stuart, Javier M. Romero, Shahmir Kamalian, Ramon G. Gonzalez, Michael H. Lev, Synho Do, An explainable deep-learning algorithm for the detection of acute intracranial hemorrhage from small datasets; 2018 Springer Nature, Nature Biomedical Engineering 3 (2019) 173–182.
 M.R. Arbabshirani, B.K. Fornwalt, G.J. Mongelluzzo, Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration, Springer Nature, 2018.