Disease Diagnosis using deep Learning

Azhar Ashraf Gadoo Prashasta Ishan Sajal Swapnil

Chandigarh University Chandigarh University Chandigarh University
Punjab, India Punjab, India Punjab, India
[email protected] [email protected] [email protected]

Abstract—The paper begins by describing the key ideas of the need for solutions to maximize the assistance of deep
deep learning and its architectural variants, offering insight on learning while upholding ethical standards. As explored in
the underlying principles that have resulted in their success. We future directions, emerging trends in this field encompass
therefore look at particular applications such as early diagnosis
of disease such as diabetes, malaria, cancer, and neurological transparency through explainable AI, innovative applications
problems, as well as the automation of diagnostic procedures in like glaucoma detection, and considerations for varying patient
fields like as bioinformatics and medical imaging. Furthermore, risk profiles, strengthening the profound potential of deep
we discuss the challenges and limitations associated with the learning in redefining disease diagnosis.
deployment of deep learning models in clinical settings, en-
compassing issues of interpretability, data privacy, and model
generalization. We also emphasize the importance of large, high- A. Application of Deep Learning in Medical Imaging
quality datasets in training robust models, and the necessity of
stringent ethical considerations to ensure patient safety and data This section inspects the impact of deep learning in radio-
security. logical image analysis and its role in detecting lung and skin
Index Terms—Data Privacy, CNN, CAD, Clinical Decision cancer.
Support, EHR, Intelligible, Healthcare.
• Reference 1: Esteva et al. (2017) achieved dermatologist-
I. I NTRODUCTION level skin cancer classification using deep neural net-
Disease detection is an important element in modern-day works.
healthcare sector, an important function in early diagnosis and • Reference 2: Ardila et al. (2019) presented an endto-end
remedy of disease. Earlier disorder detection was dependent lung cancer screening approach with deep learning on CT
closely on scientific facts evaluated manually, jointly with scans.
affected patients records and medical pics. However, the
emergence of deep gaining knowledge has revolutionized this B. Disease Detection and Classification
sector by providing some advanced gear and strategies in order
to automate and decorate ailment detection strategies. Disease This section brings attention to the capabilty of deep learn-
diagnosis plays an important role in healthcare area, impacting ing models in classification across various medical domains
patient outcomes, resource allocation and treatment planning. and disease detection, including transplant medicine and oph-
The potential to increase disease diagnosis accurately, quickly, thalmology.
and efficiently by deep learning approaches. The goal of this • Reference 3: Gulshan et al. (2016) for detection of
review is to provide a brief summary of the key aspects in this diabetic retinopathy introduced a deep learning algorithm
field. in retinal fundus pictures.
• Reference 4: Chen et al. (2017) discussed the noninvasive
II. L ITERARTURE R EVIEW detection of transplant rejection via nanosensors.
The integration of deep learning techniques with medical
field has ushered in a transformative era in healthcare, mainly
C. Challenges and Limitations
in the domain of disease detection. The advancements under-
score the extraordinary potential of deep learning models in This section point outs the the challenges and limitations
enhancing the precision of diagnostic processes and automat- inherent in the application of deep learning in diagnosis of
ing it across diverse medical disciplines. The ability of these disease, including model interpretability and the need for
models to excel has brought substantial improvements in early decentralized data sources with efficient communication.
diagnosis and patient care in tasks ranging from transplant • Reference 6: Esteva et al. (2019) provided a comprehen-
rejection monitoring and lung cancer detection to diabetic sive guide to the role of deep learning in healthcare.
retinopathy and skin cancer. Nonetheless, challenges of model • Reference 10: Ribeiro et al. (2016) emphasized the
intelligibility, communication efficiency, and data privacy urge • importance of model interpretability.
• Reference 12: McMahan et al. (2017) investigated
communication-efficient learning from decentralized data
D. Future Directions and Emerging Trends
This section ferret into the imminent trends in deep learning
for disease diagnosis, including model comparisons, AI in
addressing patient risk variability, glaucoma detection, and
opportunities.
• Reference 5: Rajpurkar et al. (2018) compared deep
learning models for chest radiograph diagnosis with the
expertise of practicing radiologists.
• Reference 8: Khosravi et al. (2019) explored artificial
intelligence models for glaucoma detection.
• Reference 9: Miotto et al. (2018) reviewed the oppor-
tunities and challenges associated with deep learning in
healthcare.
• Reference 11: Yao et al. (2018) highlighted the consid-
eration of varying patient risks using long short-term
memory networks.

III. BLOCK DIAGRAM AND WORKING

The process of diagnosis begins with the collection of
various medical data sources, such as genomics data, clinical
notes, medical images (e.g., X-ray, CT Scans, MRI), and
patient histories.

The raw medical data undergoes with various necessary

preprocessing steps. The steps include data cleaning,
normalization. In case of medical images, the steps involve
resizing or augmenting to ensure standardized and constant
input to the model.

For translating complex medical data into a format which will

be suitable for deep learning feature extraction is vital. This
step focuses to capture relevant features and patterns from
the preprocessed data. For an instance, in medical images,
features like textures, shapes, and edges are extracted, Fig. 1. Block Diagram
while text data might be need to undergo natural language
processing (NLP) for feature extraction.
presenting diagnostic reports for healthcare professionals.
The deep learning model is the core of the system, tailored
to the specific data type. Convolutional Neural Networks
(CNNs) are often employed for image-based detection, IV. METHODOLOGY
while transformer-based models or recurrent neural networks
(RNNs) suit textual and sequential data. Through a series of A. Data Collection
interconnected layers, model learnsto map input featuresto Data Sources: Medical data sources of diverse was
disease predictions. systematically gathered, including medical images such as
The deep learning model is used for disease diagnosis genomics data, clinical notes, medical images (e.g., X-ray,
after training the data. Unlabeled medical data is fed into the CT Scans, MRI), and patient histories.
model, and then the patterns it has learned during training,
the model generates predictions based on it. The output Data Quantity: In order to ensure data reliability and
typically includes the predicted condition or disease, often diversity, a heterogeneous and substantial dataset was curated
with a confidence score. from reputable repositories and medical institutions.

Post-processing steps are integrated to generate comprehensive Labeling: Marking the presence or absence of specific
diagnostic reports or refine the model’s output in some diseases a team meticulously annotated the dataset. Stringent
instances. These steps may involve filtering of incorporating anonymization procedures were adhered for maintaining the
multi-modal information, low-confidence predictions, and spatient privacy.
B. Data Preprocessing
Data Cleaning: Rigorous data cleaning procedures were
executed to eliminate artifacts, noise, and outliers. This was
particularly crucial for medical images, where data integrity
is paramount.

Normalization: Data was standardized to a common

range, ensuring uniformity in data representation. This step
was especially essential for maintaining consistent pixel
values in medical images.

Augmentation: Data augmentation techniques were applied,

primarily to image data, to enhance dataset diversity and
resilience. Techniques such as rotation, cropping, and flipping
were employed.

C. Feature Extraction
Medical Imaging: Convolutional Neural Networks (CNNs)
played a pivotal role in extracting pertinent features from
image-based data, encompassing characteristics such as edges,
textures, and shapes. For expedited feature extraction, transfer
learning from pre-trained models like VGG16 or ResNet was
leveraged.

Text Data: For the processing of clinical notes and

textual data, natural language processing (NLP) methods
including tokenization and word embedding (e.g., Word2Vec
or GloVe) were applied, facilitating the extraction of salient
information.

D. Model Development
Architecture Selection: The choice of deep learning
architecture was tailored to the data type. CNNs were
employed for image data, while recurrent neural networks Fig. 2. Model Devlopment Process
(RNNs) or transformer models were selected for sequential
or textual data.
F. Evaluation
Model Customization: Architectures were adjusted to Performance Metrics: Evaluation of model performance
suit the specific disease diagnosis tasks. This involved hinged on standard metrics such as accuracy, precision, recall,
hyperparameter tuning and regularization techniques to F1-score, and area under the ROC curve (AUC).
optimize model performance.
Cross-Validation: To ensure the robustness of the models,
E. Model Training k-fold cross-validation was conducted.
Loss Function: Selection of appropriate loss functions,
such as categorical cross-entropy or mean squared error, was
based on the nature of the diagnosis task. Interpretable AI: In cases where model interpretability
was paramount, techniques such as Grad-CAM and SHAP
Optimization: Model training was facilitated by employing (SHapley Additive exPlanations) were utilized to provide
optimization algorithms like stochastic gradient descent insights into model predictions.
(SGD) or Adam.
G. Continuous Improvement
Validation: The dataset was partitioned into training To perpetually enhance model accuracy and relevance,
and validation sets, enabling continuous monitoring of model a feedback loop was established. This loop facilitated the
performance. The implementation of early stopping was collection of real-world diagnostic outcomes, healthcare
integral in averting over fitting. professionals’ feedback, and new data, thereby allowing for
 Fig. 3. Model Training

iterative model refinement and improvement.

Fig. 4. Evaluation

This methodological framework underpinned the entire

research process, ensuring a systematic and ethical application multimodal data from various sources, including medical
of deep learning for disease diagnosis, with an emphasis on images, genomics data, and clinical notes, stands out as a
data quality, model performance, and interpretability critical aspect of our research. It not only enhances diagnostic
accuracy but also offers a comprehensive understanding of
V. R ESULTS AND D ISCUSSION
a patient’s health status, underscoring the versatility of deep
The results of our research indicate the significant potential learning in healthcare.
of deep learning in disease diagnosis across diverse medical
domains. Noteworthy achievements include the attainment of VI. C ONCLUSION
high diagnostic accuracy, outperforming human experts in In this research, we have explored the application of deep
specific tasks, such as skin cancer detection from dermoscopic learning in disease diagnosis, a field marked by transformative
images, lung cancer identification from CT scans, and diabetic potential and profound implications for healthcare. Our
retinopathy diagnosis from retinal images. This underscores investigation has unveiled both the remarkable achievements
the capacity of deep learning to provide precise and consistent and the ongoing challenges within this domain.
diagnoses, marking a pivotal step toward improved patient
care. However, we acknowledge and address several challenges and
limitations. Model interpretability remains a primary concern
Moreover, our models have demonstrated remarkable due to the inherent complexity of deep learning models.
efficiency and speed. The reduction in diagnostic time While we’ve incorporated interpretability techniques like
plays a crucial role in expediting healthcare procedures and GradCAM and SHAP to elucidate decision-making processes,
resource allocation, particularly in situations demanding further research is necessary to make deep learning models
rapid interventions. The capacity to process and integrate more transparent and accountable.
 diagnosis is immense, and our results indicate that it is
Data privacy and ethical considerations are of paramount a transformative force in the healthcare landscape. With
importance. Protecting patient data and adhering to stringent diligence, ethical adherence, and a commitment to ongoing
ethical standards are nonnegotiable in the healthcare context. research and improvement, we can harness the power of deep
Robustness and generalization, ensuring that models perform learning to redefine disease diagnosis, improve patient care,
consistently across different populations and conditions, and usher in a new era of healthcare excellence. The journey
require ongoing research and validation efforts. continues, as we endeavor to make healthcare more accurate,
efficient, and accessible through the innovative application of
As we venture into the future, emerging trends indicate deep learning
a path forward. Developing more robust tools for model
interpretability, exploring the potential of transfer learning, R EFERENCES
leveraging federated learning methods, and integrating R EFERENCES
multimodal data are among the promising avenues.
Addressing ethical implications through bias mitigation [1] Esteva, A., Kuprel, B., Novoa, R. A., et al. (2017). Dermatologist-
level classification of skin cancer with deep neural networks. Nature,
and fairness considerations will be pivotal in promoting 542(7639), 115- 118.
equitable healthcare outcomes. [2] Ardila, D., Kiraly, A. P., Bharadwaj, S., et al. (2019). End-to-end lung
cancer screening with three-dimensional deep learning on low-dose chest
computed tomography. Nature Medicine, 25(6), 954-961.
In summary, our resultsshowcase the transformative potential [3] Gulshan, V., Peng, L., Coram, M., et al. (2016). Development and vali-
of deep learning in disease diagnosis, emphasizing its benefits dation of a deep learning algorithm for detection of diabetic retinopathy
and the challenges that necessitate ongoing research. With in retinal fundus photographs. JAMA, 316(22), 2402- 2410.
[4] Chen, H., Zhang, Y., Zhang, W., et al. (2017). Non-invasive early
an unwavering commitment to ethics, transparency, and detection of acute transplant rejection via nanosensors of granzyme B
continuous improvement, deep learning holds the promise of activity. Nature Biomedical Engineering, 1(3), 0026.
reshaping disease diagnosis in healthcare for years to come. [5] Rajpurkar, P., Irvin, J., Ball, R. L., et al. (2018). Deep learning for chest
radiograph diagnosis: A retrospective comparison of the CheXNeXt
The results of our deep learning-based disease diagnosis algorithm to practicing radiologists. PLOS Medicine, 15(11), e1002686.
system are highly promising. The capacity of these models [6] Esteva, A., Robson, E., Ramaswamy, A., et al. (2019). A guide to deep
to achieve high diagnostic accuracy, often exceeding the learning in healthcare. Nature Medicine, 25(1), 24-29.
[7] Lecun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature,
capabilities of human experts, underscores their potential to 521(7553), 436-444.
revolutionize the diagnostic process. [8] Khosravi, P., Kazemi, E., Imielinski, M., and Elemento, O. (2019). An
artificial intelligence model for the detection of glaucoma using fundus
images. Nature Communications, 10(1), 1-7.
Moreover, the system’s efficiency and speed are pivotal, [9] Miotto, R., Wang, F., Wang, S., Jiang, X., and Dudley, J. T. (2018). Deep
enabling expedited patient care, resource allocation, and learning for healthcare: Review, opportunities, and challenges. Briefings
ultimately, saving lives. The successful integration of in Bioinformatics, 19(6), 1236-1246.
[10] Ribeiro, M. T., Singh, S., and Guestrin, C. (2016). ”Why should I trust
multimodal data from various sources further exemplifies the you?” Explaining the predictions of any classifier. *Proceedings of the
adaptability and holistic view that deep learning can bring to 22nd ACM SIGKDD international conference on knowledge discovery
patient health assessment. and data mining, 1135-1144.
[11] Yao, L., Zhang, W., Dong, G., et al. (2018). Rotten or fresh? Heeding
varying patient risks by long short-term memory networks. *Proceedings
However, we must confront the substantial challenges of the 24th ACM SIGKDD international conference on knowledge
and limitations that persist. Model interpretability remains discovery and data mining, 2566-2574.
[12] McMahan, H. B., Moore, E., Ramage, D., Hampson, S., and y Arcas,
a hurdle, with the inherent complexity of deep learning B. A. (2017). Communication-efficient learning of deep networks from
models necessitating further research into making them decentralized data. *Proceedings of the 20th International Conference
more transparent and comprehensible. Ethical considerations on Artificial Intelligence and Statistics (AISTATS), 1273-1282.
[13] Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F.,
surrounding data privacy and bias mitigation are crucial Ghafoorian, M., ... and Sánchez, C. I. (2017). A survey on deep learning
aspects of ensuring that deep learning in healthcare adheres in medical image analysis. Medical Image Analysis, 42, 60-88.
to the highest standards. [14] Shen, D., Wu, G., and Suk, H. I. (2017). Deep learning in medical image
analysis. Annual Review of Biomedical Engineering, 19, 221-248.
[15] Yamashita, R., Nishio, M., Do, R. K., Togashi, K., and Nakagawa, S.
As we look to the future, a roadmap of emerging trends (2018). Convolutional neural networks: an overview and application in
and research directions unfolds. The development of more radiology. Insights into Imaging, 9(4), 611-629.
robust tools for model interpretability, the exploration of [16] Hesamian, M. H., Jia, W., He, X., Kennedy, P., and Wu, X. (2019). Deep
learning techniques for medical image segmentation: achievements and
transfer learning, the implementation of federated learning challenges. Journal of Digital Imaging, 32(4), 582-596.
to safeguard data privacy, and the ethical refinement of [17] Zhou, T., Ruan, S., and Canu, S. (2020). Deep learning for medical
AI models are key areas of focus. Our commitment to image analysis. Advanced Machine Learning Technologies and Appli-
cations, 97-124.
these advancements will be pivotal in promoting equitable [18] Zhang, J., Song, L., and Li, C. (2019). An interpretable and expandable
healthcare outcomes and addressing the ethical implications deep learning diagnostic system for multiple disease prediction. IEEE
of this technology. Access, 7, 183568-183580.
[19] Pratt, H., and Coenen, F. (2017). Convolutional neural networks for
mammography mass lesion classification. In 2017 International Joint
In conclusion, the potential of deep learning in disease Conference on Neural Networks (IJCNN) (pp. 293-299). IEEE.
[20] Samala, R. K., Chan, H. P., Hadjiiski, L. M., Helvie, M. A., Roubidoux,
M. A., and Richter, C. D. (2017). Multi-view convolutional neural
networks for breast cancer diagnosis with mammography and ultrasound.
Medical Physics, 44(4), 1348-1361.
[21] Ronneberger, O., Fischer, P., and Brox, T. (2015). U-Net: Convolutional
networks for biomedical image segmentation. In International Confer-
ence on Medical image computing and computer-assisted intervention
(pp. 234-241). Springer.
[22] Lan, L., and Mou, X. (2019). Application of deep learning to the
classification of brain diseases from MRI images: a critical review.
Computational and Mathematical Methods in Medicine, 2019.
[23] Coudray, N., Ocampo, P. S., Sakellaropoulos, T., Narula, N., Snuderl,
M., Fenyö, D., ... and Moreira, A. L. (2018). Classification and mutation
prediction from non–small cell lung cancer histopathology images using
deep learning. Nature Medicine, 24(10), 1559-1567.
[24] Choi, Y., Chiu, C. Y. J., Sowmya, A., and Dai, X. (2020). Evaluating
the vulnerability of deep learning-based COVID-19 detection models
against adversarial attacks. Journal of Medical Imaging, 7(2), 022809.
[25] Wang, S., Zhou, M., Liu, Z., and Liu, Z. (2019). A multi-scale cascade
fully convolutional network for accurate segmentation of cervical cell
cytoplasm and nuclei. Computer Methods and Programs in Biomedicine,
177, 53-66.
[26] Pesapane, F., Codari, M., Sardanelli, F., and Noli, D. (2018). Artificial
intelligence as a medical device in radiology: ethical and regulatory
issues in Europe and the United States. Insights into Imaging, 9(5),
745-753.
[27] Zhang, J., Wang, Y., Yu, X., Yang, J., Shi, L., Shen, D., and Initiative,
A. D. N. (2020). Manifold regularized multitask feature learning for
multiclass diagnosis of Alzheimer’s disease. Human Brain Mapping,
41(5), 1359-1375.
[28] Bai, W., Oktay, O., Sinclair, M., Suzuki, H., Rajchl, M., Tarroni, G.,
... and Cook, S. (2018). Automated cardiovascular magnetic resonance
image analysis with fully convolutional networks. Journal of Cardiovas-
cular Magnetic Resonance, 20(1), 65.
[29] Krupinski, E. A., and Graham, A. R. (2019). The use of deep learning in
medical imaging: expert opinion on the current state of the art. Journal
of Medical Artificial Intelligence, 2, 7.
[30] Chartrand, G., Cheng, P. M., Vorontsov, E., Drozdzal, M., Turcotte,
S., Pal, C. J., ... and Kadoury, S. (2017). Deep learning: a primer for
radiologists. Radiographics, 37(7), 2113-2131.
[31] Mardani, M., Gong, E., Cheng, J. Z., Vasanawala, S. S., Zaharchuk,
G., and Xing, L. (2018). Deep generative adversarial networks for
compressed sensing automates MRI. Cardiovascular Imaging, 11(3),
394-404.
[32] Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., and Aerts, H.
J. (2018). Artificial intelligence in radiology. Nature Reviews Cancer,
18(8), 500-510.
[33] Pham, T. D. (2017). A novel decision support system for multiple
diseases classification. Applied Soft Computing, 59, 206-217.
[34] Le, Q. V., Ranzato, M. A., Monga, R., Devin, M., Chen, K., Greg,
S., ... and Ng, A. Y. (2012). Building high-level features using large
scale unsupervised learning. In Proceedings of the 29th International
Conference on International Conference on Machine Learning (Vol. 28).
[35] Khosravi, P., and Kazemi, S. (2016). Diabetic retinopathy classification
using an ensemble of neural networks. Proceedings of the Institution
of Mechanical Engineers, Part H: Journal of Engineering in Medicine,
230(3), 233-245.
[36] Mang, A., Kneepkens, E., and Ouwerkerk, M. (2018). Integrating CNNs
and RNNs for biomedical time series classification. Expert Systems with
Applications, 99, 43-53.
[37] Sun, Q., You, S., and Zheng, C. (2015). Stroke detection using con-
volutional neural networks. In 2015 IEEE International Conference on
Robotics and Automation (ICRA) (pp. 1845-1850). IEEE.