Artificial intelligence diagnosis system for nervous system diseases

Abstract

The invention relates to the technical field of digital medical treatment, and particularly discloses an artificial intelligence diagnosis system for nervous system diseases. The system comprises a data acquisition module, a preprocessing module, a multi-modal feature extraction module, a feature fusion module, a disease classification module, an interpretability analysis module and a doctor feedback and model updating module. The system supports acquisition and fusion of medical images, electroencephalogram signals, biochemical indexes, questionnaire information and wearable device data, constructs unified feature representation through a weighted attention mechanism, and outputs a prediction result of nervous system diseases based on a deep learning model. And meanwhile, the system is combined with a Shapley value method to realize the interpretability of the diagnosis process, and supports closed-loop return of doctor feedback information and model increment updating. The nervous system disease diagnosis accuracy and early recognition capability are improved, and the method has relatively high clinical deployability and adaptive optimization capability.

Description

Technical Field

[0001] This invention relates to the field of digital medical technology, and more specifically, to an artificial intelligence diagnostic system for neurological diseases. Background Technology

[0002] Neurological diseases are among the major types of diseases affecting human health and quality of life, including Parkinson's disease, Alzheimer's disease, epilepsy, post-stroke sequelae, and many other chronic neurodegenerative and functional disorders. These diseases often manifest as symptoms such as memory loss, motor impairment, language disorders, and epileptic seizures. The disease progresses slowly and varies significantly from person to person, making early diagnosis difficult, with a high rate of misdiagnosis and frequent delays in intervention.

[0003] Traditional clinical diagnosis of neurological diseases relies primarily on neuroimaging examinations (such as MRI and CT), electroencephalography (EEG), neuropsychological assessments (such as MMSE and MoCA), laboratory biochemical tests, and the physician's clinical experience. While these methods are widely used in clinical practice, they have several significant problems: First, the data from various examinations are scattered across different departments or equipment, requiring physicians to manually integrate them, which is inefficient and prone to missing information; second, diagnostic results are highly dependent on the physician's experience level, and there may be significant differences in judgment between different physicians; third, some patients do not show obvious changes on imaging or have normal questionnaire scores in the early stages of the disease, making early identification and intervention impossible.

[0004] In recent years, artificial intelligence technology, especially deep learning, has made groundbreaking progress in fields such as medical imaging, speech recognition, and text understanding, providing new solutions for improving the auxiliary diagnostic capabilities of neurological diseases. Existing research has attempted to apply convolutional neural networks (CNNs) to brain MRI image analysis, extracting features from brain atrophy or abnormal structural regions to identify Parkinson's disease or Alzheimer's disease. Other methods use recurrent neural networks (RNNs, LSTMs) to analyze changes in EEG signal rhythms or seizure patterns. However, these approaches are mostly based on single-modal data sources (such as images or EEG alone), lacking the integration of multi-source heterogeneous data, thus limiting diagnostic accuracy and generalization ability.

[0005] Furthermore, existing AI models generally suffer from the "black box" problem, failing to provide convincing diagnostic evidence. Doctors struggle to understand the pathological features upon which the model bases its judgments, hindering the widespread adoption of these systems in clinical practice. Simultaneously, the lack of a systematic data update mechanism makes it difficult for models to adapt to changes brought about by regional differences, equipment variations, or the evolution of disease patterns.

[0006] Therefore, this application designs and develops an artificial intelligence diagnostic system for nervous system diseases to solve the above-mentioned problems. Summary of the Invention

[0007] The purpose of this invention is to solve the technical problems mentioned in the background section and to provide an artificial intelligence diagnostic system for nervous system diseases.

[0008] The above-mentioned objective of the present invention is achieved as follows:

[0009] An artificial intelligence diagnostic system for neurological diseases, characterized in that it comprises:

[0010] Data acquisition module: Used to acquire multimodal input data, including medical image data (D). img EEG data D eeg Text data D text Biochemical index data D bio and wearable device physiological data D wear ;

[0011] Data preprocessing module: Performs format conversion, noise removal, and normalization on the input data to obtain preprocessed medical image data X. img EEG data X eeg Text data X text Biochemical indicator data X bio and wearable device physiological data X wear , denoted as {X img ,X eeg ,X text ,X bio ,X wear};

[0012] Multimodal feature extraction module: Extracts feature vectors f for each modality respectively. img ,f eeg ,f text ,f bio ,f wear ;

[0013] Feature fusion module: Calculates a unified feature vector F using a weighted attention mechanism;

[0014] Disease classification module: Outputs a predicted probability vector P for nervous system diseases based on the fused feature vector F;

[0015] Interpretability analysis module: Based on the Shapley value mechanism, it evaluates the contribution of each modality to the prediction results and outputs a structured diagnostic report through the front-end interactive interface;

[0016] Doctor feedback module: Used to receive doctor correction tags And update the model parameters through an incremental optimization mechanism;

[0017] The model update module is used to incrementally learn and update the diagnostic model parameters θ based on doctor feedback.

[0018] As a preferred embodiment of the present invention, the image feature extraction module is constructed based on a convolutional neural network and outputs an image feature vector f. img Expressed as:

[0019] f img =φ(W c *X img +b c );

[0020] in, Indicates input of medical images, W c For the convolution kernel, b c φ represents the bias term, * indicates the convolution operation, and φ(·) represents the nonlinear activation function, preferably the ReLU function.

[0021] As a preferred technical solution of the present invention, the EEG feature extraction module is based on a bidirectional long short-term memory network and outputs a feature vector f. eeg Expressed as:

[0022]

[0023] in, These represent the hidden states of the Bi-LSTM at the T-th time step in both the forward and reverse directions. This represents vector concatenation, where d is the dimension of the LSTM unidirectional hidden layer.

[0024] As a preferred technical solution of the present invention, the text data feature extraction module is based on the Transformer structure and outputs a feature vector f. text Expressed as:

[0025] f text =SelfAttention(X) text )·W t ;

[0026] in, Let l represent a text input embedding of length l, and SelfAttention(·) represent the self-attention mechanism. This is the weight matrix.

[0027] As a preferred technical solution of the present invention, the multimodal feature fusion module uses a cross-attention mechanism to construct a unified feature vector F, which is expressed as:

[0028]

[0029]

[0030] in, For each modal feature vector, Let α be the attention projection vector. i ∈[0,1] represents the modal weight coefficients, satisfying ∑α i =1.

[0031] As a preferred technical solution of the present invention, the disease classification module is based on a fully connected layer and a softmax function, and outputs a probability distribution vector P for each neurological disease:

[0032] P = Softmax(W o ·F+b o );

[0033] in, k represents the number of categories. in This indicates the diagnostic result.

[0034] As a preferred technical solution of the present invention, the interpretability analysis module evaluates the contribution of each modal feature based on the Shapley value method, and its calculation formula is as follows:

[0035]

[0036] Among them, SHAP i Let v(S) be the explanatory score of the i-th feature, representing the estimation function of the feature subset S for the predicted value, and |F| be the total number of features.

[0037] As a preferred embodiment of the present invention, the doctor feedback module is used to receive feedback tags. The current model parameters are then optimized based on the following loss function:

[0038]

[0039] Among them, y i For real labels, p i The output probability is Softmax, where θ is the current model parameter. ref Here are the parameters of the model from the previous iteration, and λ is the coefficient of the regularization term.

[0040] As a preferred technical solution of the present invention, the front-end interactive interface outputs a structured diagnostic report, including: predicted disease labels. Modal feature weights α i Feature Contribution Map (SHAP) i Abnormal indicator alerts and suggestions for further inspection.

[0041] As a preferred technical solution of the present invention, the system supports an incremental training optimization mechanism for the model, and its parameter update adopts the following strategy:

[0042]

[0043] Where, θ t Let X be the model parameters for the t-th round, η be the learning rate, and X be the model parameters for the t-th round. new y new For the newly added samples and labels, L is the aforementioned loss function.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. This invention introduces a multimodal collaborative diagnostic mechanism, comprehensively utilizing medical images, electroencephalogram (EEG) signals, biochemical indicators, structured questionnaires, and wearable device data to construct a unified feature fusion and judgment model. Compared to traditional diagnostic methods that rely on single images or EEG data, this system effectively improves the ability to discriminate complex and heterogeneous neurological diseases, especially demonstrating higher recognition sensitivity and specificity in the early stages of Parkinson's disease and Alzheimer's disease, thereby significantly improving the accuracy and prospectiveness of clinical diagnosis.

[0046] 2. The system of this invention employs a weighted attention fusion network and an interpretability analysis module. It can not only dynamically identify key diagnostic features of different modalities in different cases, but also present the basis for model decisions to doctors through mechanisms such as Shapley values, achieving transparency and visualization of the "black box" algorithm. Doctors can obtain clear diagnostic support information, enhancing their trust in the system's conclusions, improving the efficiency of human-machine collaborative judgment, and providing structured, traceable intelligent suggestions for complex cases.

[0047] 3. The solution of this invention constructs a closed-loop feedback mechanism, supporting doctors to correct diagnostic results in real time and return labels. The system can perform small-batch incremental learning based on new samples, continuously optimizing model parameters to adapt to changes in the characteristics of different hospitals, equipment, or populations, and has good cross-scenario adaptability and continuous evolution capabilities. This system is not only suitable for large medical institutions, but can also be deployed in community clinics, telemedicine platforms, and other scenarios, realizing the clinical feasibility and widespread application of the technology. Attached Figure Description

[0048] Figure 1 This is a system block diagram of an artificial intelligence diagnostic system for nervous system diseases according to an embodiment of the present invention;

[0049] Figure 2 This is a logic flowchart of an artificial intelligence diagnostic system for nervous system diseases according to an embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] The implementation of the present invention will be described in detail below with reference to specific embodiments.

[0052] Example 1: See Figure 1 and Figure 2 As shown in the figure, the embodiment of the present invention provides an artificial intelligence diagnostic system for nervous system diseases, which can integrate multimodal data such as medical images, electroencephalogram signals, biochemical indicators, text question answering and wearable devices to perform high-precision and interpretable auxiliary diagnosis of nervous system-related diseases (including but not limited to Alzheimer's disease, Parkinson's disease, epilepsy, etc.), and continuously optimize model performance based on doctor feedback to achieve AI diagnostic services that combine personalization and clinical deployability.

[0053] The system described in this application first acquires the patient's multimodal input data through a data acquisition module. The acquired data includes five categories:

[0054] First, medical imaging data D img The images are mainly derived from brain MRI or CT scans.

[0055] Second, EEG data D eeg It can be continuously recorded and uploaded by a head-mounted EEG device;

[0056] Thirdly, questionnaire and medical record text data D text This includes scoring scales such as MMSE, ADL, and MoCA, as well as medical history information;

[0057] Fourth, biochemical indicator data D bio Such as the concentration of Aβ42 and Tau protein in the blood; fifth, physiological data from wearable devices. wear This includes heart rate variability (HRV), gait, sleep parameters, etc.

[0058] The collected data will be uniformly sent to the data preprocessing module for standardization processing, specifically including image data size normalization and contrast enhancement; bandpass filtering and artifact removal of EEG signals; word segmentation and word vector conversion of text; and time alignment and numerical normalization of all modal data, ultimately forming the processed standard data input set {X}. img ,X eeg ,X text ,X bio ,X wear}

[0059] The data from each modality are then input into the feature extraction module, image data X. img Spatial structure features are extracted using a deep convolutional neural network (CNN), and these features are expressed as follows:

[0060] f img =φ(W c *X img +b c );

[0061] in, Indicates the input image, W c For the convolution kernel, b c The term represents the bias term, * indicates the convolution operation, φ(·) is the non-linear activation function (ReLU), and the output feature vector is... EEG Data X eeg The data is fed into a Bi-LSTM neural network model, and its temporal features are extracted to obtain a concatenated bidirectional hidden state vector.

[0062]

[0063] in, These represent the final hidden states of the Bi-LSTM network in the forward and reverse directions, respectively. This indicates a splicing operation to obtain the final EEG feature vector. Text input After self-attention and weight mapping, a semantic feature vector is generated:

[0064] f text =SelfAttention(X) text )·W t ;

[0065] in, It is the parameter matrix obtained through training, and the output is the text feature vector.

[0066] Biochemical indicators and wearable device data were also extracted into fixed-dimensional vectors using lightweight fully connected networks. and To ensure dimensional consistency during subsequent fusion.

[0067] Subsequently, all modal feature vectors will be input into the feature fusion module, where they will be fused into a unified diagnostic feature vector F through a weighted attention mechanism.

[0068]

[0069] Among them, f i ∈{f img ,f eeg ,ftext ,f bio ,f wear} are all feature vectors of dimension d; For the attention vector obtained during training, α i The weight coefficients for this modality reflect its diagnostic importance in the current sample. The fused unified feature vector F enters the disease classification module, and the probability distribution of each disease category is output through the Softmax activation function:

[0070] P = Softmax(W o ·F+b o );

[0071] in, k represents the total number of categories of diseases to be diagnosed. The output vector P = p1, p2, ..., p k This represents the predicted probability for each category, with the system using the maximum index as the final predicted label:

[0072]

[0073] To enhance the interpretability of the system, a Shapley value mechanism is further introduced to quantify the contribution of each modality or feature subset in the prediction process:

[0074]

[0075] Among them, SHAP i Representing feature f i The marginal contribution value of the final prediction, v(S), is the output estimate of the feature subset S in the model. This interpretation mechanism can not only provide doctors with diagnostic basis, but also reveal the specific source of abnormal feature combinations.

[0076] The system also supports a doctor feedback mechanism. If a doctor finds that the model has made a misjudgment, they can provide feedback via a feedback tag. To correct predictions and trigger model fine-tuning, the loss function incorporates a regularization term as follows:

[0077]

[0078] Among them, y i For real labels, p i For the corresponding predicted probability, θ is the current parameter, θ ref This is a snapshot of the parameters from the previous iteration, where λ is the regularization coefficient used to suppress overfitting. The updated parameters can be iterated using incremental gradient optimization.

[0079]

[0080] Where η is the learning rate, X new ,ynew This mechanism provides doctors with additional feedback data, ensuring the model can continuously learn and evolve with real clinical data, exhibiting excellent personalization and generalization capabilities.

[0081] Finally, the system displays a structured diagnostic report on the front-end diagnostic interface, including: disease prediction. Modal feature weights α i Explanation of the SHAP graph i The system can detect abnormal patterns, suggest further examinations (such as PET imaging or genetic testing), and recommend treatment pathways. It also supports access to remote consultation services, allowing structured reports to be shared with experts for online review and collaborative decision-making.

[0082] In summary, this system integrates multimodal data, multi-layer neural network structure, attention mechanism and feedback loop to build an AI diagnostic system for neurological diseases that is practical, highly accurate, and has strong interpretability. It is especially suitable for intelligent auxiliary decision-making scenarios in primary healthcare institutions, specialized hospitals and telemedicine platforms.

[0083] Example 2: In the neurology outpatient department of a large tertiary hospital, a 67-year-old male patient presented with tremor and bradykinesia in his right hand. The initial physical examination failed to determine the nature of the disease. He had no history of significant brain injury and had not received any treatment related to Parkinson's disease. In order to assist doctors in making an earlier and more accurate diagnosis, the patient was included in the artificial intelligence diagnostic system for neurological diseases proposed in this invention for intelligent assisted diagnosis.

[0084] The system first collects the patient's multimodal health data through a combination of the hospital's information platform and wearable devices. This includes T1-weighted MRI images of the head, 16-channel EEG signals from a 5-minute resting state, standardized questionnaire completion results (MMSE score 27, UPDRS motor score 18), blood biochemistry test results (α-synuclein concentration 1.9 ng / mL, IL-6 concentration 3.1 pg / mL), and physiological indicators such as average gait speed (1.1 m / s) and decreased heart rate variability collected via a smartwatch over the past week. These data undergo unified preprocessing by the system, including image size standardization, signal denoising, wavelet transform, text embedding encoding, and normalization, forming a standardized set of five input features: Image X img EEG X eeg Text X text Biochemical X bio and wearable X wear After modular feature extraction, each modality is transformed into a 128-dimensional numerical feature vector. The image data is then fed into a five-layer convolutional neural network, where features are extracted using the following formula:

[0085] fimg =φ(W c *X img +b c );

[0086] Where the convolution kernel parameter W c Bias b c The activation function φ(·) was learned during the system training phase. The EEG data was input into the Bidirectional Long Short-Term Memory (Bi-LSTM) network, and its forward and reverse terminal states were concatenated as follows:

[0087]

[0088] After extracting contextual representations from text data using a self-attention mechanism, they are then compared with the parameter matrix W. t Multiplication yields the text embedding vector:

[0089] f text =SelfAttention(X) text )·W t ;

[0090] Meanwhile, biochemical and wearable data, after being structured and mapped, are extracted into... At this point, all five modal feature vectors are of the same dimension, which satisfies the requirements for subsequent fusion.

[0091] The system employs a weighted attention mechanism to fuse multimodal features into a unified diagnostic representation vector F. The attention parameter vector is set as follows:

[0092] W a = [0.10, 0.12, 0.15, 0.11, 0.13];

[0093] The system employs a weighted attention mechanism to fuse multimodal features into a unified diagnostic representation vector F. The attention parameter vector is set as follows:

[0094] W a = [0.10, 0.12, 0.15, 0.11, 0.13];

[0095] For each mode f i Perform the following attention weight scoring and normalization operations:

[0096]

[0097] Taking simulated data calculation as an example, the attention scores for each modality are as follows: image 0.2263, EEG 0.2593, text 0.1803, biochemistry 0.1723, and wearable 0.1515. After Softmax normalization, the modal weights assigned by the system are: image 0.2057, EEG 0.2125, text 0.1964, biochemistry 0.1948, and wearable 0.1906, demonstrating that EEG and image hold dominant weights in the current judgment.

[0098] The system then performs modal fusion:

[0099]

[0100] Taking the first dimension of the fusion vector as an example, the actual calculation is as follows:

[0101] F1=0.2057×0.22+0.2125×0.47+0.1964×0.25+0.1948×0.33+0.1906×0.20≈0.2967.

[0102] Similarly, the complete fused feature vector is calculated.

[0103] The system feeds the fused features into a multi-class Softmax neural network classifier to determine the disease category. After linear transformation:

[0104] z = W o ·F+b o = [2.34, 1.82, 0.85, 0.13];

[0105] After executing the Softmax function, the output probability vector is:

[0106] P=[0.5182,0.3081,0.1168,0.0569];

[0107] The first category is Parkinson's disease, with a confidence level of 51.82%, and the system ultimately outputs the diagnostic label as "Parkinson's disease".

[0108] To improve interpretability, the system further performed Shapley value analysis on the contribution of each modality feature. The calculation results showed that the signal abnormality in the substantia nigra region of the MRI image contributed the most to the judgment (SHAP = +0.30), followed by the enhancement of slow waves in the EEG (+0.24), the high score of the UPDRS questionnaire (+0.15), the increase of α-synuclein (+0.10), and gait changes (+0.05).

[0109] These results were used to generate a structured diagnostic report in the form of charts and text, and displayed to clinicians through a front-end interface. After analyzing the judgments and interpretations provided by the system, the doctors believed that the conclusions were reasonable to a certain extent. However, considering the patient's history of cerebrovascular disease, they suggested adding "possible vascular Parkinson's disease" as a differential diagnosis. The doctors then corrected the labels through the system interface. The system immediately records the feedback and triggers the model fine-tuning module. The model is trained and updated using the following loss function:

[0110]

[0111] The system uses the current parameter θ t Learning rate η, new samples (X) new ,y new Perform one round of gradient descent:

[0112]

[0113] The optimized model is immediately deployed and updated, and the sample is included in the subsequent model training database. The entire process, from data collection, algorithm inference, result output, interpretation and display to doctor feedback and model correction, forms a closed-loop intelligent diagnosis and treatment process.

[0114] This embodiment demonstrates that the artificial intelligence diagnostic system for neurological diseases of the present invention not only possesses the technical characteristics of being accurate, interpretable, and deployable, but also has practical application value in clinical practice, and is particularly suitable for the auxiliary identification and risk screening of neurological diseases such as Parkinson's disease and Alzheimer's disease.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence diagnostic system for nervous system diseases, characterized in that, include: Data acquisition module: Used to acquire multimodal input data, including medical image data (D). img EEG data D eeg Text data D text Biochemical index data D bio and wearable device physiological data D wear ; Data preprocessing module: Performs format conversion, noise removal, and normalization on the input data to obtain preprocessed medical image data X. img EEG data X eeg Text data X text Biochemical indicator data X bio and wearable device physiological data X wear , denoted as {X img X eeg X text X bio X wear }; Multimodal feature extraction module: Extracts feature vectors f for each modality respectively. img f eeg f text f wear ; Feature fusion module: Calculates a unified feature vector F using a weighted attention mechanism; Disease classification module: Outputs a predicted probability vector P for nervous system diseases based on the fused feature vector F; Interpretability analysis module: Based on the Shapley value mechanism, it evaluates the contribution of each modality to the prediction results and outputs a structured diagnostic report through the front-end interactive interface; Doctor feedback module: Used to receive doctor correction tags And update the model parameters through an incremental optimization mechanism; The model update module is used to incrementally learn and update the diagnostic model parameters θ based on doctor feedback.

2. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The image feature extraction module is built on a convolutional neural network and outputs an image feature vector f. img Expressed as: f img =φ(W c *X img +b c ); in, Indicates input of medical images, W c For the convolution kernel, b c The term is the bias term, * represents the convolution operation, and φ(·) represents the nonlinear activation function.

3. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The EEG feature extraction module is based on a bidirectional long short-term memory network and outputs a feature vector f. eeg Expressed as: in, , , respectively, represent the hidden states of the Bi-LSTM at the T-th time step in the forward and backward directions, ⊕ indicates vector concatenation, and d is the dimension of the LSTM unidirectional hidden layer.

4. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The text data feature extraction module is based on the Transformer structure and outputs a feature vector f. text Expressed as: f text =SelfAttention(X text )·W t ; in, Let l represent a text input embedding of length l, and SelfAttention(·) represent the self-attention mechanism. This is the weight matrix.

5. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The multimodal feature fusion module uses a cross-attention mechanism to construct a unified feature vector F, denoted as: in, For each modal feature vector, Let α be the attention projection vector. i ∈[0,1] represents the modal weight coefficients, satisfying ∑α i =1.

6. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The disease classification module is based on a fully connected layer and a softmax function, outputting the probability distribution vector P for each neurological disease. P=Softmax(W o ·F+b o ); in, k represents the number of categories. P = [p1, p2, ..., p k ],in This indicates the diagnostic result.

7. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The interpretability analysis module evaluates the contribution of each modal feature based on the Shapley value method. The calculation formula is as follows: Among them, SHAP i Let v(S) be the explanatory score of the i-th feature, representing the estimation function of the feature subset S for the predicted value, and |F| be the total number of features.

8. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The doctor feedback module is used to receive feedback tags. The current model parameters are then optimized based on the following loss function: Among them, y i For real labels, p i The output probability is Softmax, where θ is the current model parameter. ref Here are the parameters of the model from the previous iteration, and λ is the coefficient of the regularization term.

9. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The front-end interactive interface outputs a structured diagnostic report, including: predicted disease labels. Modal feature weights α i Feature Contribution Map (SHAP) i Abnormal indicator alerts and suggestions for further inspection.

10. The artificial intelligence diagnostic system for nervous system diseases according to claim 1, characterized in that, The model update module supports an incremental training optimization mechanism, and its parameter updates employ the following strategy: Where, θ t Let X be the model parameters for the t-th round, η be the learning rate, and X be the model parameters for the t new y new For the newly added samples and labels, L is the aforementioned loss function.

Images