A parkinson's disease early detection method based on efficientnetv2-s

Abstract

This invention proposes an early detection method for Parkinson's disease based on EfficientNetV2-S. Addressing the problems of traditional diagnosis relying on subjective assessment of motor symptoms by doctors, which is prone to misdiagnosis and missed diagnosis in the early stages, this method acquires handwritten images through spiral drawing and waveform handwriting tasks. It then utilizes EfficientNetV2-S to automatically learn subtle jitter features and motion control anomalies in the trajectory. Multi-scale feature fusion enhances the perception of minute local changes, and dual-task joint learning improves the comprehensive discrimination ability across different handwriting tasks. After automatic alignment, size normalization, and data augmentation, the data is input into the model, achieving high-precision, high-efficiency, non-invasive, and easily scalable early screening and assisted diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis and intelligent assisted diagnosis technology, specifically involving a method for early detection of Parkinson's disease based on deep learning, and in particular a method for automated, objective, and early identification of Parkinson's disease by utilizing spiral drawing tasks and waveform handwriting tasks, combined with the EfficientNetV2-S network structure. Background Technology

[0002] Parkinson's disease (PD) is a common neurodegenerative disease of the central nervous system. Its main clinical manifestations include resting tremor, muscle rigidity, bradykinesia, and postural instability. Because the symptoms of PD are insidious in the early stages and vary significantly from person to person, traditional clinical diagnostic methods rely heavily on the experience of neurologists and are easily influenced by subjective factors, leading to misdiagnosis or missed diagnosis.

[0003] Existing auxiliary diagnostic methods, such as brain imaging analysis, biochemical testing, or long-term behavioral observation, generally suffer from high testing costs, strong equipment dependence, complex operation, and unsuitability for large-scale screening. Therefore, exploring a low-cost, non-invasive, easily deployable Parkinson's disease detection technology suitable for early screening is of significant clinical importance.

[0004] Previous studies have shown that Parkinson's disease patients exhibit significant differences in handwriting compared to healthy individuals when completing the Spiral Drawing Task and Wave Drawing Task, particularly in terms of tremor amplitude, line smoothness, speed variation, and trajectory deviation. These handwriting tasks can indirectly reflect a patient's fine motor control ability, thus serving as important behavioral biomarkers for early screening of Parkinson's disease.

[0005] With the rapid development of deep learning technology in the field of computer vision, convolutional neural networks have demonstrated excellent performance in medical image analysis tasks. However, traditional deep networks still face the following problems in the analysis of handwritten images related to Parkinson's disease:

[0006] 1) The model has a large number of parameters, resulting in low inference efficiency, which is not conducive to practical deployment;

[0007] 2) Insufficient ability to model fine-grained local jitter features;

[0008] 3) Limited feature generalization ability for multi-task handwriting modes (spiral and waveform).

[0009] Therefore, it is necessary to design an improved deep learning detection method that balances high accuracy and high efficiency and can effectively model the subtle motion features of handwriting tasks. Summary of the Invention

[0010] 1. To address the above problems, this invention proposes an early detection method for Parkinson's disease based on EfficientNetV2-S, comprising the following steps:

[0011] Step 1: Construct the input data acquisition method for the spiral drawing task and the waveform handwriting task, and form standardized input samples.

[0012] Step 2: Propose a handwritten image preprocessing method to improve input consistency and robustness through automatic alignment, cropping and normalization adjustment.

[0013] Step 3: Propose a feature extraction model with EfficientNetV2-S as the backbone network to extract deep discriminative features of handwritten trajectories.

[0014] Step 4: Propose a multi-branch joint learning and multi-scale feature fusion module to classify and discriminate spiral tasks, waveform tasks and fused features respectively.

[0015] Step 5: Propose a multi-loss joint optimization function. By weighting the spiral branch loss, waveform branch loss, fusion classification loss and regularization term, the stability and generalization performance of the model are improved, enabling early detection of Parkinson's disease.

[0016] 2. The data acquisition in step 1 further includes the following features:

[0017] The spiral drawing task involves the subject drawing a spiral curve in a preset template or blank area to obtain a spiral handwritten image; then the waveform handwriting task involves the subject writing a waveform curve along a preset path to obtain a waveform handwritten image; in addition, the acquisition devices include, but are not limited to, digital handwriting tablets, tablet computers, smartphone touch screens or scanning devices, and the acquired data can be offline images or raster images after online trajectory conversion;

[0018] Secondly, a sample index is established for each subject, including at least one sample of the spiral task and one sample of the waveform task, to ensure the consistency of the dual task inputs. Finally, the samples are divided into training set, validation set and test set, and the category labels are uniformly coded, with the labels including at least the Parkinson's disease category and the healthy control category.

[0019] 3. The preprocessing method in step 2 specifically includes the following:

[0020] First, the input handwritten image is converted to grayscale and denoised, and threshold segmentation or adaptive binarization is used to enhance handwriting contrast. Second, the principal orientation of the image is determined through edge detection, Hough transform, or principal orientation estimation, and automatic rotation correction is performed to ensure that the handwritten content is consistent with the preset standard orientation. Furthermore, the handwriting region is located and cropped using minimum bounding rectangle or connected component analysis to reduce background interference. The ROI image is scaled to the model input size while maintaining the aspect ratio, and padding is applied when necessary. Pixel normalization is then performed. Finally, data augmentation strategies are employed to improve generalization ability, including rotation, scaling, translation, slight deformation perturbation, and noise perturbation.

[0021] 4. The EfficientNetV2-S feature extraction model in step 3 further includes the following features:

[0022] First, the EfficientNetV2-S staged convolutional structure is used to extract local texture and global morphological features from the handwritten image step by step. Second, the network includes a fusion convolutional module and a moving inverse residual module, which improves the efficiency of feature representation through depthwise separable convolution and channel attention mechanism. Then, global average pooling is introduced at the end of the backbone network to compress the two-dimensional feature map into a one-dimensional feature vector for classification. Finally, the feature vector is mapped to the target class space through a linear layer or classification head to output the corresponding classification probability.

[0023] 5. The multi-branch joint learning and multi-scale feature fusion module in step 4 specifically includes the following:

[0024] First, multi-scale feature maps are extracted from different stages of EfficientNetV2-S to simultaneously preserve local jitter details and overall structural information. Then, 1×1 convolutions are applied to the multi-scale feature maps to align the number of channels, and feature fusion is achieved through upsampling and concatenation to form a fused feature representation. Next, at least three classification branches are constructed: a spiral branch, a waveform branch, and a fusion branch. The spiral and waveform branches classify the corresponding task features, and the fusion branch performs the final discrimination of the fused features. Finally, the output of the fusion branch is used as the main discrimination result, and consistency constraints or weighted integration can be applied to the outputs of the spiral and waveform branches to improve stability.

[0025] 6. The multi-loss joint optimization function in step 5 specifically includes the following:

[0026] The total loss function is composed of the spiral branch classification loss, waveform branch classification loss, fusion branch classification loss, and a weighted regularization term; and the expression for the total loss function is:

[0027]

[0028] in, It is a spiral branch classification loss. It is the number of grid cells. It is the number of bounding boxes predicted for each grid cell. For regularization terms; These are the weighting coefficients.

[0029] The classification loss mentioned above uses cross-entropy loss or weighted cross-entropy loss, and overfitting can be suppressed and the model's generalization ability improved through L2 regularization, Dropout, or early stopping strategies.

[0030] Compared with the existing technologies, the present invention has the following main technical advantages:

[0031] 1) Low cost and easy to promote: Early assisted screening can be completed solely based on handwritten task data, making it suitable for grassroots and home settings;

[0032] 2) High efficiency and strong deployability: With EfficientNetV2-S as the backbone network, it balances high accuracy with low computational overhead, making it easy to deploy on mobile or edge devices.

[0033] 3) Enhanced robustness: Automatic alignment and standardized preprocessing reduce the impact of different acquisition conditions and improve generalization ability;

[0034] 4) Dual-task complementarity: Joint learning of spiral and waveform tasks enhances the ability to represent subtle motion anomalies (jitter, speed instability, trajectory deviation) and improves early detection rate and stability. Attached Figure Description

[0035] Figure 1 This is a diagram of an EfficientNet network structure provided by the present invention;

[0036] Figure 2 This is the dataset provided by the present invention;

[0037] Figure 3 This invention provides the training accuracy curve of EfficientNet (SPIRAL).

[0038] Figure 4 This invention provides the EfficientNet (SPIRAL) confusion matrix diagram.

[0039] Figure 5 This invention provides the EfficientNet (WAVE) confusion matrix diagram.

[0040] Figure 6 This invention provides a training accuracy curve for EfficientNet (WAVE).

[0041] Figure 7 This invention provides a bar chart comparing the performance indicators of multiple models. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0043] Example 1: In this example, the dataset consists of spiral drawing task images and waveform handwriting task images. Acquisition methods include, but are not limited to, digital handwriting tablets, tablet touch capture, paper scanning, etc. Each subject includes at least one set of spiral task samples and one set of waveform task samples, labeled with the category label "Parkinson's Disease (PD)" or "Healthy Control (HC)".

[0044] In this embodiment, the evaluation metrics used include: accuracy, precision, recall, and F1 score, and ROC-AUC can be used as an auxiliary evaluation metric.

[0045] Accuracy measures how well a model correctly predicts all samples; it is the proportion of samples correctly predicted by the model out of the total number of samples. It reflects the overall accuracy of the model's classification. The formula for calculating accuracy is as follows:

[0046]

[0047] In this formula, TP represents the number of samples that the model correctly identifies as positive examples, i.e., the number of prohibited items correctly identified by the model; while FP represents the number of samples that the model incorrectly identifies as positive examples, i.e., the number of non-prohibited items that the model incorrectly identifies as prohibited items.

[0048] Precision measures the proportion of samples that the model predicts as positive (PD in this example) but which are actually positive (PD). It reflects the model's reliability in predicting PD, i.e., the degree of false positives (classifying HC as PD). The precision formula is as follows:

[0049]

[0050] In this formula, TP represents the number of samples that the model correctly identifies as PD; FP represents the number of samples that the model incorrectly identifies as PD.

[0051] Recall measures the proportion of samples that are actually positive (PD) that are successfully identified as PD by the model. It reflects the model's ability to detect PDs, i.e., the degree of false negatives (classifying PDs as HCs). The formula for recall is as follows:

[0052]

[0053] In this formula, TP represents the number of samples that the model correctly identifies as PD; FN represents the number of samples that the model incorrectly identifies as HC.

[0054] The F1 score is used to comprehensively evaluate precision and recall; it is the harmonic mean of the two. When it is necessary to consider both false positives and false negatives, the F1 score can more comprehensively reflect the model performance. The formula for calculating the F1 score is as follows:

[0055]

[0056] Precision and Recall represent precision and recall, respectively.

[0057] The ROC curve is used to depict the performance change of the model under different classification thresholds. The horizontal axis represents the false positive rate (FPR), and the vertical axis represents the true positive rate (TPR). AUC is the area under the ROC curve, which measures the model's overall ability to distinguish between positive and negative samples. The larger the AUC value, the stronger the model's ability to distinguish between them.

[0058]

[0059]

[0060] TPR represents the true positive rate (consistent with the recall rate); FPR represents the false positive rate.

[0061] These evaluation metrics collectively provide a comprehensive quantitative basis for model performance. Accuracy reflects the overall correctness of the model's predictions across all samples, directly demonstrating the overall classification effect. Precision focuses on the proportion of samples predicted as PDs that are actually PDs, measuring the reliability of the model's PD predictions and reflecting the degree of false positives (misclassifying HCs as PDs). Recall focuses on the proportion of samples that are actually PDs that are correctly identified as PDs, measuring the model's ability to detect PDs and reflecting the degree of false negatives (misclassifying PDs as HCs). The F1 score combines precision and recall, providing a more balanced evaluation between false positives and false negatives, especially suitable for scenarios with uneven class distribution or where both recognition accuracy and detection rate need to be considered simultaneously. Furthermore, ROC-AUC can be used as an auxiliary metric to supplement the evaluation of classification performance from the perspective of the model's overall discriminative ability at different thresholds. In summary, these metrics can reveal the model's strengths and weaknesses from multiple dimensions, providing important basis for model performance analysis and subsequent optimization.

[0062] The experimental environment for this embodiment is as follows: Processor: AMD R9 7945Hx; Memory: 32GB DDR5 (5200MHz); Graphics card: NVIDIA GeForce RTX 4060; Storage: 1024GB SSD; Operating system: Windows 11.

[0063] Development environment: Python 3.9; PyTorch (2.8.0+cu126); torchvision (0.23.0+cu126). Table 1: Algorithm Performance Comparison Evaluation indicators Precision Recll ROC-AUC This article discusses EfficientNet (SPIRAL). 88% 75% 85% EfficientNet (WAVE) 90% 76% 88% Resnet50 76% 75% 80%

[0064] As can be seen from the comparison, with the support of dual-task joint learning and fusion branch discrimination, the present invention can make fuller use of the complementary information between the spiral task and the waveform task, and has higher sensitivity to early motion control anomalies, thus having an advantage over the baseline model in terms of accuracy, recall and F1 score.

[0065] In summary, this invention proposes an early detection method for Parkinson's disease based on EfficientNetV2-S, aiming to address the problems of traditional early screening methods, such as reliance on human experience, high subjectivity, and high equipment costs. Through joint modeling of spiral drawing and waveform handwriting tasks, and by employing preprocessing standardization, multi-branch fusion discrimination, and multi-loss joint optimization strategies, this invention achieves efficient, stable, and deployable assisted screening for Parkinson's disease, possessing significant practical application value, particularly suitable for scenarios such as primary care screening, remote health management, and mobile initial screening.

[0066] The above embodiments illustrate only one implementation of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for early detection of Parkinson's disease based on EfficientNetV2-S, characterized in that, The specific steps are as follows: Step 1: Collect handwriting task data from the subjects, wherein the handwriting task includes at least spiral drawing task images and waveform writing task images; Step 2: Construct a feature extraction network with EfficientNetV2-S as the backbone to perform deep feature extraction on the spiral drawing task image and the waveform writing task image; Step 3: Introduce a multi-scale feature fusion module to enhance the ability to express fine-grained features such as local handwriting jitter, line thickness variation, and trajectory offset; Step 4: Establish a dual-task joint learning structure of spiral branch and waveform branch, and fuse the features of different tasks to obtain a fused feature vector; Step 5: Construct a classification loss function for early detection of Parkinson's disease, perform classification prediction on the fused feature vector, and output the probability results of Parkinson's disease / healthy control to achieve early detection of Parkinson's disease.

2. The method for early detection of Parkinson's disease based on EfficientNetV2-S according to claim 1, characterized in that, The EfficientNetV2-S backbone network in step 2 further includes the following features: The input spiral and waveform images are fed into EfficientNetV2-S, respectively. Edge, texture, and structural features of the handwriting are extracted through convolutional layers and non-linear activation layers. The backbone network employs Fused-MBConv and MBConv structures to achieve efficient feature extraction, reducing parameter count and inference overhead while maintaining accuracy. Feature maps output from different stages of the backbone network are used as a multi-scale feature set, providing high-resolution detail and low-resolution semantic information for subsequent multi-scale fusion. A global feature aggregation operation is introduced at the end of the backbone to compress the two-dimensional feature map into a one-dimensional feature vector for subsequent branch discrimination and fusion.

3. The method for early detection of Parkinson's disease based on EfficientNetV2-S according to claim 1, characterized in that, The multi-scale feature fusion module in step 3 specifically includes the following: Feature maps are extracted from at least two different layers of the backbone network, the feature maps containing high-resolution detail features and low-resolution semantic features; the feature maps at different scales are aligned, including upsampling / downsampling to make the spatial size consistent; the aligned features are fused by concatenation or weighted fusion, and channel compression and feature reshaping are performed through convolutional layers; Output the fused multi-scale feature map to improve the detection capabilities for local jitter textures, subtle irregular fluctuations, and overall trajectory shape changes.

4. The method for early detection of Parkinson's disease based on EfficientNetV2-S according to claim 1, characterized in that, The dual-task joint learning structure in step 4 specifically includes the following: First, a shared backbone feature extraction layer is set up, which extracts shared representations from the spiral image and the waveform image; second, a spiral task branch is set up to learn the continuity, smoothness, radius change stability and jitter characteristics of the spiral trajectory. Then, a waveform task branch is set up to learn waveform periodicity, amplitude fluctuation, line jitter, and rhythm instability features. Finally, the feature vectors output by the two branches are fused. The fusion method includes feature concatenation, weighted summation, or attention-weighted fusion to obtain a fused feature vector for final classification.

5. The method for early detection of Parkinson's disease based on EfficientNetV2-S according to claim 1, characterized in that, The process between S1 and S4 includes an image preprocessing step, which specifically includes the following: First, the original handwritten image is automatically oriented to obtain a standardized image with a uniform writing direction. Second, the image is normalized to the required size for model input while maintaining the aspect ratio, and padding is used when necessary. Furthermore, the image is grayscaled and intensity normalized to reduce domain differences caused by different acquisition devices and lighting conditions. Finally, data augmentation is performed, including at least rotation, scaling, translation, slight noise perturbation, and random erasure, to improve the model's generalization ability.

6. The method for early detection of Parkinson's disease based on EfficientNetV2-S according to claim 1, characterized in that, The classification loss function in step 5 specifically includes the following: First, cross-entropy loss is used as the basic classification loss to achieve binary classification learning between Parkinson's disease patients and healthy controls. When the sample classes are imbalanced, Focal Loss or class-weighted cross-entropy is introduced to balance the contributions of positive and negative samples. The total loss function consists of branch loss and fusion loss, satisfying: in, For spiral branching classification loss, For waveform branch classification loss, To fuse feature classification loss, For regularization terms; These are the weighting coefficients.

Images