Detection of Parkinson’s Disease Using Convolutional Neural Networks

Detection of Parkinsons Disease Using Convolutional Neural Networks

Dr. Nandha Gopal S M1, Ankitha K D2, Bhumika K P3, Chaithanya K4, and Deepika Lakshmi S5

1Professor and Head, Dept. of CSE, ACS College of Engineering, ACSCE, Bangalore, India 21AH22CS014, Dept. of CSE, ACS College of Engineering, ACSCE, Bangalore, India 31AH22CS025, Dept. of CSE, ACS College of Engineering, ACSCE, Bangalore, India 41AH22CS030, Dept. of CSE, ACS College of Engineering, ACSCE, Bangalore, India 51AH22CS044, Dept. of CSE, ACS College of Engineering, ACSCE, Bangalore, India

AbstractParkinsons Disease (PD) is a progressive neurological disorder that primarily affects the motor system, leading to symptoms such as tremors, rigidity, and impaired movement. Accurate early diagnosis remains challenging because PD shares clinical characteristics with several other neurological conditions, contributing to an estimated 25% inaccuracy rate in manual diagnosis. This paper presents a Convo- lutional Neural Network (CNN)-based automated system designed to classify PD patients from healthy controls (HC) using T2-weighted Magnetic Resonance Imaging (MRI) data sourced from the Parkinsons Progression Markers Initiative (PPMI). Mid-brain slices from 500 MRI scans are selected and spatially aligned via image registration. The system pipeline includes grayscale conversion, noise removal, feature extraction, and CNN-based classification. Performance evaluation using accuracy, sensitivity, specificity, and AUC demonstrates that the pro- posed CNN model outperforms existing techniques by 3%9% across all metrics. The system provides a reliable, automated decision-support tool to assist neurologists in early and consistent diagnosis.

KeywordsParkinsons Disease; Convolutional Neural Network; Deep Learning; MRI; Medical Image Analysis; Early Diagnosis

1. INTRODUCTION

   Parkinsons Disease is a neurodegenerative disorder affecting millions of individuals worldwide. It predominantly impacts the motor system, producing hallmark symptoms such as rest- ing tremors, bradykinesia, muscle rigidity, and postural insta- bility. Early-stage diagnosis is essential because timely in- tervention significantly slows symptom progression and im- proves quality of life for patients.

   Conventional diagnosis relies on clinical observation and neu- rological assessments performed by specialists. This ap- proach is inherently subjective and often fails to detect the disease during its early phase, when physical manifestations are subtle and overlap with other conditions. Pereira et al. reported that hand-drawn exams of both healthy individuals and early-stage PD patients appear similar, highlighting the difficulty of early detection

   Advances in deep learning, particularly Convolutional Neu- ral Networks, have demonstrated strong performance across a range of medical image classification tasks

   This research proposes a CNN-based automated system that processes T2-weighted MRI images to classify subjects as ei- ther PD-affected or healthy. The system integrates prepro- cessing, feature extraction, and deep learning-based classifi- cation into a unified pipeline, with the aim of providing a con- sistent and scalable diagnostic aid.

2. PROBLEM DEFINITION

   Detecting Parkinsons Disease, especially at an early stage,

   involves several practical difficulties:

   • Early-stage ambiguity: Motor symptoms in PDs initial stages closely resemble those of other neurological con- ditions. Hand-drawn diagnostic exams from healthy and early-stage PD patients are often indistinguishable

   • Limited and biased datasets: MRI datasets specific to PD

     graphic diversity of the patient population, making general- isation of trained models a persistent challenge.

   • Cost of at-home monitoring: Continuous patient monitor- ing outside clinical settings requires specialised and expen- sive hardware, limiting long-term follow-up

   • Subjectivity in manual diagnosis: Clinician-based assess- ment introduces variability, contributing to the approxi- mately 25% misdiagnosis rate reported in the literature.

     An automated system that can reliably analyse MRI scans and distinguish PD-affected patients from healthy individuals ad- dresses each of these challenges by reducing subjectivity and scaling to large datasets.

3. EXISTING SYSTEM

   Current approaches to Parkinsons Disease diagnosis include: Clinical Observation: Neurologists assess motor symptoms such as tremors and gait disturbances. While practical, this method is subjective and dependent on the clinicians experi- ence.

   Manual MRI Analysis: Radiologists inspect brain scans for structural abnormalities. This process is time-consuming and prone to inter-observer variability, particularly for subtle early-stage changes.

   Traditional Machine Learning: Earlier computational ap- proaches applied Support Vector Machines (SVM), K-Nearest Neighbours (KNN), and Naive Bayes classifiers to engineered feature sets. These methods require domain-specific feature extraction and do not generalise well to varied imaging con- ditions

   Imaging Modalities: Positron Emission Tomography (PET) and Single Photon Emission Computed Tomography (SPECT) offer valuable biomarker information but involve ra- dioactive tracers and higher costs, limiting their routine use. Structural MRI has historically played a limited role; how-

   potential for detecting PD-related brain changes

4. PROPOSED SYSTEM

   The proposed system applies a deep CNN architecture to T2- weighted MRI images to automate the classification of Parkinsons Disease. Key advantages of this approach in- clude:

   Automated Feature Learning: The CNN learns discrimina- tive spatial features from raw MRI data through stacked con- volutional and pooling layers, removing the need for manual feature engineering.

   Contextual Spatial Understanding: Convolutional opera- tions capture local neighbourhood relationships in brain im- ages, enabling the model to detect subtle structural changes associated with PD that are imperceptible to the human eye. Standardised Pipeline: The system applies a consistent pre- processing routineincluding noise removal, grayscale con- version, and image registrationensuring uniformity across different patient scans before classification.

   Quantitative Evaluation: Model outputs are evaluated using accuracy, sensitivity, specificity, and the Area Under the ROC Curve (AUC), providing a comprehensive picture of diagnos- tic performance beyond simple accuracy metrics.

5. OBJECTIVE

   The specific objectives of this project are:

   • To collect T2-weighted MRI brain images of PD patients and healthy individuals from the PPMI public dataset.

   • To preprocess MRI data through grayscale conversion, me- dian filtering for noise removal, and image registration for spatial alignment of mid-brain slices.

   • To implement a CNN-based classification model that dis- tinguishes PD-affected subjects from healthy controls.

   • To evaluate model performance using accuracy, sensitivity, specificity, and AUC metrics.

   • To compare the deep learning models performance against traditional machine learning approaches.

6. LITERATURE SURVEY

   A growing body of research has applied machine learning and deep learning technques to Parkinsons Disease detec- tion across multiple data modalities.

   Bhan and Kapoor proposed a deep learning model for classi- fying PD from MRI scans, demonstrating that automated ap- proaches substantially reduce diagnostic time while improv- ing accuracy over manual methods

   Qiu et al. introduced a multiscale convolutional prototype net- work for PD detection using EEG signals, leveraging multi- scale feature extraction to achieve robust cross-subject gener- alisation

   Rizvi et al. proposed UEN-PDNet, a deep neural network for classifying PD from resting-state EEG, capturing discrimina- tive temporal-spatial signal patterns with strong accuracy Zhang et al. analysed EEG-based brain functional networks for early-stage PD detection, constructing connectivity graphs and applying machine learning classifiers to network-level features

   Dai et al. conducted a comprehensive review of data-driven PD diagnostic systems across gait, voice, and imaging modal-

   isons

   Demir et al. presented an LSTM-based framework that maps speech features to capture temporal dysarthric patterns, demonstrating that voice recordings serve as reliable non- invasive PD biomarkers

   Talitckii et al. identified optimal wearable sensor exercises for PD detection using machine learning, contributing to evidence-based data-collection protocol design

   Chauhan and Ghosal introduced a hybrid CNN-BiLSTM model that combines spatial feature extraction with tempo- ral sequence learning on multimodal inputs including gait, speech, and handwriting data

   Santos et al. converted EEG data into graph representations and applied residual neural networks for PD classification, capturing spatial-topological electrode relationships Malekroodi et al. utilised large self-supervised speech models fine-tuned for PD detection, representing a recent shift toward foundation-model-based medical diagnostics

7. SOFTWARE REQUIREMENTS

   • Operating System: Windows 10/11 or Linux

   • Programming Language: Python 3.8 or higher

   • Libraries: OpenCV, TensorFlow/Keras, NumPy, Pandas, Matplotlib, Scikit-learn, Seaborn

   • IDE: Jupyter Notebook / Visual Studio Code

   • Toolbox: Image Processing Toolbox (OpenCV)

8. HARDWARE REQUIREMENTS

   • Processor: Intel Core i3/i5 at 2.4 GHz or better

   • RAM: 4 GB minimum (8 GB recommended)

   • Storage: 500 GB hard disk

   • GPU: NVIDIA GPU recommended for model training; CPU sufficient for inference

9. METHODOLOGY

   The proposed methodology is structured as a sequential image processing and classification pipeline:

   Image Collection: T2-weighted MRI scans of 500 subjects are obtained from the PPMI public dataset, comprising both PD-affected patients and healthy controls.

   Image Preprocessing:

   • Grayscale Conversion: RGB MRI images are converted to 8-bit grayscale, reducing memory requirements by approx- imately 33% and simplifying downstream feature computa- tion.

   • Noise Removal: A non-linear median filter is applied us- ing a sliding window of odd length. Each sample within the window is sorted by magnitude and replaced by the me- dian value, preserving edge information while suppressing random pixel noise.

   • Image Enhancement: Global thresholding removes back- ground regions and retains structures of diagnostic interest. A high-pass filter amplifies fine spatial details, sharpening structural boundaries within the mid-brain.

     Image Segmentation: The processed image is partitioned into meaningful regions using a mean-shift clustering algo- rithm. This sliding-window approach iteratively converges toward high-density pixel clusters, isolating brain structures from surrounding tissue.

     (HOG) descriptors are computed from preprocessed images. For each pixel, the horizontal and vertical gradient compo- nents (Gx and Gy) are calculated, from which gradient mag- nitude and orientation angle are derived. Orientation frequen- cies are binned into histograms that constitute the feature vec- tor fed into the classifier.

     CNN-Based Classification: The preprocessed image is passed directly through the CNN, which performs its own in- ternal feature learning via stacked convolutional, rectification, and pooling layers. A SoftMax output layer produces class probability scores for the PD and HC categories.

     Evaluation: Model performance is assessed using accuracy, sensitivity, specificity, and AUC to provide a multidimen- sional view of diagnostic reliability.

10. SYSTEM DESIGN

    The overall system architecture processes MRI data through a sequence of well-defined modules:

    1. Image Collection acquisition of T2-weighted MRI scans.

    2. Image Preprocessing grayscale conversion, noise re- moval, and image enhancement.

    3. Image Segmentation region-of-interest extraction using mean-shift clustering.

    4. Feature Extraction HOG descriptor computation.

    5. Training CNN model training on labelled MRI slices.

    6. Classification prediction of PD or healthy status.

       1. CNN Architecture

          The network consists of alternating convolutional and pooling layers followed by fully connected layers and a SoftMax out- put. The convolutional layer applies learnable filters across the input image to generate feature maps. The pooling layer performs max-pooling to downsample spatial dimensions and retain dominant activations. The fully connected layer flattens the feature maps into a single vector, which is then mapped to class probabilities by the output layer

       2. MVC Architecture

    The application follows a Model-View-Controller pattern. The Model encapsulates the CNN inference logic. The View is the Python Tkinter GUI, which accepts MRI image input from the user and displays the classification result. The Con- troller coordinates data flow between the interface and the deep learning backend.

11. SYSTEM TESTING

    Testing was carried out on a Windows 10 machine with an Intel Core i5 processor and 8 GB RAM, using Python with OpenCV and TensorFlow libraries. The GUI was validated using the Python Tkinter framework.

    Unit Testing: Individual functionsincluding grayscale con- version, median filtering, and CNN inferencewere tested in isolation to verify that each module produces correct outputs independently. This phase ensured that foundational compo- nents were defect-free before system integration.

    Integration Testing: The communication between the Tkin- ter frontend and the CNN backend was validated to confirm that selected MRI images are correctly forwarded to the in- ference engine and that classification results are returned and displayed without format mismatches. Both bottom-up and

    top-down integration strategies were applied.

    System Testing: The complete application was evaluated as a unified system against all specified functional and non- functional requirements. Testing verified OS compatibility across Windows XP and Windows 10, confirming that perfor- mance is best on Windows 10. Latency measurements con- firmed that the system delivers predictions within an accept- able response time for practical clinical use.

12. RESULT

    Experimental evaluation demonstrates that the CNN-based mdel outperforms traditional machine learning classifiers across all metrics. The system accurately distinguishes PD- affected MRI scans from healthy controls, particularly in cases where early-stage structural differences are subtle.

    Table 1. Model Performance Comparison

    Method	Accuracy (%)	AUC
    SVM (Traditional)	87	0.85
    KNN	83	0.81
    CNN (Proposed)	95	0.97

    The proposed CNN model achieved an improvement of 3% 9% over existing techniques in terms of accuracy, sensitivity, specificity, and AUC, consistent with findings reported in the literature for deep learning-based PD detection

13. CONCLUSION

This paper presented a CNN-based automated system for the early detection of Parkinsons Disease using T2-weighted MRI brain images. The proposed pipeline integrates grayscale conversion, median filtering, global thresholding, HOG-based feature extraction, and deep CNN classification into a cohesive diagnostic framework.

Experimental results confirm that the CNN model sig- nificantly outperforms conventional machine learning ap- proaches by capturing spatial feature hierarchies that man- ual methods miss. The system reduces diagnostic subjectiv- ity, scales to large imaging datasets, and provides quantitative confidence scores to support clinical decision-making.

Future work will focus on training the model on larger and more diverse patient cohorts to improve generalisation. In- tegration of complementary modalitiesincluding EEG sig- nals, speech biomarkers, and gait datacan further enhance diagnostic robustness. Incorporating explainable AI tech- niques such as Grad-CAM will improve clinical interpretabil- ity, and deployment as a cloud-based web application will support remote patient monitoring.

ACKNOWLEDGMENT

The authors would like to thank the faculty and staff of the Department of Computer Science and Engineering, ACS Col- lege of Engineering, Bangalore, for their guidance and sup- port throughout this work. The authors also acknowledge the Parkinsons Progression Markers Initiative (PPMI) for provid- ing the publicly available MRI dataset used in this research.

REFERENCES

1. S. Shetty and Y. S. Rao, SVM based machine learning approach to iden- tify Parkinsons disease using gait analysis, in Proc. Int. Conf. on Inventive Computation Technologies (ICICT), vol. 2, IEEE, 2016.

2. P. P. Michel, E. C. Hirsch, and S. Hunot, Understanding dopaminergic cell death pathways in Parkinson disease, Neuron, vol. 90, no. 4, pp. 675691, 2016.

3. G. Pahuja and T. N. Nagabhushan, A novel GA-ELM approach for Parkinsons disease detection using brain structural T1-weighted MRI data, in Proc. 2nd Int. Conf. on Cognitive Computing and Information Processing (CCIP), IEEE, 2016.

4. S. J. Chinta and J. K. Andersen, Dopaminergic neurons, The Int. J. of Biochemistry & Cell Biology, vol. 37, no. 5, pp. 942946, 2005.

5. S. Little, A. Pogosyan, S. Neal, B. Zavala, L. Zrinzo, M. Hariz, T. Foltynie, P. Limousin, K. Ashkan, J. FitzGerald, A. L. Green, T. Z. Aziz, and P. Brown, Adaptive deep brain stimulation in advanced Parkinson disease, Annals of Neurology, vol. 74, no. 3, pp. 449457, 2013.

6. L. M. Shulman, R. L. Taback, J. Bean, and W. J. Weiner, Comorbid- ity of the nonmotor symptoms of Parkinsons disease, Parkinsonism & Related Disorders, vol. 8, no. 3, pp. 193197, 2002.

7. J. A. Obeso, M. C. Rodriguez-Oroz, M. M. Goetz, C. Marin, J. H. Ko- rdower, M. Rodriguez, E. C. Hirsch, M. Farrer, A. H. V. Schapira, and

   G. Halliday, Missing pieces in the Parkinsons disease puzzle, Nature Medicine, vol. 16, no. 6, pp. 653661, 2010.

8. C. R. Pereira, D. R. Pereira, F. A. Silva, C. Hook, S. A. T. Weber, and

   J. P. Papa, A step towards the automated diagnosis of Parkinsons dis- ease: Analyzing handwriting movements, Pattern Recognition Letters, Elsevier, 2016.

9. G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoo- rian, J. A. W. M. van der Laak, B. van Ginneken, and C. I. Sanchez, A survey on deep learning in medical image analysis, Medical Image Analysis, Elsevier, 2017.

10. Y. LeCun, Y. Bengio, and G. Hinton, Deep learning, Nature, vol. 521, no. 7553, pp. 436444, 2015.

11. D. Shen, G. Wu, and H. I. Suk, Deep learning in medical image anal- ysis, Annual Review of Biomedical Engineering, vol. 19, pp. 221248, 2017.

12. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cam- bridge, MA: MIT Press, 2016.

13. P. Rajpurkar, J. Irvin, B. Ball, K. Zhu, B. Yang, H. Mehta, T. Duan,

    D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, M. P. Lungren, and A.

    Y. Ng, Deep learning for chest radiograph diagnosis: A retrospective

    comparison, PLOS Medicine, 2018.

14. A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, Dermatologist-level classification of skin cancer with deep neural networks, Nature, vol. 542, pp. 115118, 2017.

15. Y. Zhang, S. Wang, P. Phillips, and Z. Dong, Detection of Alzheimers disease and mild cognitive impairment based on structural volumetric MR images using 3D-CNN, IEEE Access, vol. 6, pp. 5713157143, 2018.

16. A. Bhan and S. Kapoor, Early diagnosis of Parkinsons disease in brain

    MRI using deep learning algorithm, IEEE Xplore, June 2021.

17. L. Qiu, J. Li, L. Zhong, X. Hu, and Q. Zhao, A novel EEG-based Parkinsons disease detection model using multiscale convolutional pro- totype networks, IEEE Trans. on Neural Systems and Rehabilitation Engineering, 2024.

18. S. Q. A. Rizvi, G. Wang, A. Khan, M. K. Hasan, M. A. Khan, and A. Mosavi, Classifying Parkinsons disease using resting-state EEG sig- nals and UEN-PDNet, IEEE Access, vol. 11, 2023.

19. W. Zhang, X. Han, S. Qiu, H. Zhao, and H. He, Analysis of brain functional network based on EEG signals for early-stage Parkinsons disease detection, IEEE Access, vol. 10, 2022.

20. Y. Dai, Z. Tang, Y. Wang, and Z. Xu, Data driven intelligent diagnos- tics for Parkinsons disease, IEEE Access, vol. 7, pp. 106941106953, 2019.

21. F. Demir, A. Sengur, A. Ari, K. Siddique, and M. Alswaitti, Feature mapping and deep long short-term memory network-based efficient ap- proach for Parkinsons disease diagnosis, IEEE Access, vol. 9, pp. 2471824727, 2021.

22. A. Talitckii, D. Kovalenko, M. Shcherbak, A. Mancera, A. Prefer- ansov, and P. Visser, Defining optimal exercises for efficient detection of Parkinsons disease using machine learning and wearable sensors, IEEE Trans. on Instrumentation and Measurement, vol. 70, 2021.

23. M. K. Chauhan and P. Ghosal, A hybrid CNN-BiLSTM neural network architecture for early prediction of Parkinsons disease, in Proc. IEEE Int. Symposium on Smart Electronic Systems (iSES), 2024.

24. L. O. Santos, A. G. Medeiros, P. A. L. Rego, T. Bian, D. S. Katz, and P.

    P. Reboucas Filho, Graph-based EEG analysis for Parkinsons disease

    classification: A residual neural network approach, in Proc. Int. Joint Conf. on Neural Networks (IJCNN), IEEE, 2024.

25. H. S. Malekroodi, P. Vo, G. Yi, and M. Hasan, Speech-based Parkin- sons detection using pre-trained self-supervised models, in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP), 2025.

26. M. A. Lones, S. L. Smith, J. E. Alty, S. Lliv, K. R. Dures, D. R. S. Jamieson, and A. P. Tyrrell, Evolving classifiers to recognize the move- ment characteristics of Parkinsons disease patients, IEEE Trans. on Evolutionary Computation, vol. 18, no. 4, pp. 559576, 2014.