View All Documents & Correspondence
Io T Driven Real Time Breast Tumour Identification And Classification Using Deep Learning Image Recognition Models
Abstract: The suggested invention also includes a smart multi-step inferential pipeline that can maximize diagnostic reliability in heterogeneous imaging. The framework is dynamically adjusted to changes in breast density, imaging resolution, contrast differences as well as acquisition noise through the combination of an adaptive normalization layer and domain-generalization strategy. This will make it strong in various clinical settings, such as tertiary hospitals, rural diagnostic centres and mobile screening units. In one implementation, the edge computing module runs a real-time preprocessing protocol comprising of anisotropic diffusion filtering and morphological noise removal, contrast-limited adaptive histogram equalization (CLAHE) and automated region-of-interest (ROI) detection with threshold-guided segmentation. Through local execution of the operations at the IoT node, the system limits the latency and limits the bandwidth usage, thus facilitating instant clinical feedback. The core of the deep learning uses a hybrid multi-branch convolutional structure. The first branch reveals the global structural patterns related to mass formation and tissue distortion whereas the second branch is concerned with high frequencies that represent microcalcifications and fine-grained lesion boundaries. The aggregation of these parallel feature streams is done by weighting the feature streams with an attention mechanism, which allows the model to give more weightage to diagnostically significant regions. To improve interpretability as an important condition in clinical AI systems the invention incorporates an explainable artificial intelligence (XAI) module. The module produces saliency maps, gradient-weighted activation maps, and confidence overlap maps of boundaries which are graphically used to identify suspicious areas. The above transparency facilitates the validation of clinicians and enhances the trust in the automated decision-making systems. Multi-class risk stratification, as opposed to binary classification, is done by the classification subsystem. Besides benign and malignant predictions, the system gives intermediate suspicion scores, which can be used to prioritize the triage and give early referral recommendations. The probabilistic risk estimation model can be using Bayesian calibration methods to decrease overconfidence in forecasting and enhance clinical reliability. Moreover, cloud-mediated coordination layer facilitates collaborative learning across devices by parameter aggregation which is secure. As opposed to sharing raw imaging information, only encrypted model gradients or weight updates are exchanged. This privacy-conscious architecture will meet the requirements of the medical data protection requirements and the continual enhancement of the generalization performance of the global model. The system also comprises of diagnostic confidence validation module which cross validates CNN outputs with extracted morphological descriptors like: • Tumour compactness ratio • Edge spiculation index • Variance of intensity in lesion boundary. • Measurement of fractal dimensions. • Texture entropy analysis This two-fold validation method minimizes false positives that are usually experienced on dense breast tissues and glandular structures. The invention offers a temporal comparison engine to facilitate longitudinal patient monitoring. Serial scans of the same patient on the breast can be examined to identify subtle changes in progression, rate of growth and structural changes. The temporal analysis helps oncologists to measure the tumour aggressiveness and response to treatment. In a more complex implementation, the structure allows multi-modal fusion learning, that is, mammography images may be mixed with ultrasound or MRI data. The alignment of cross-modality features enhances the diagnostic certainty especially in the unclear cases. Mathematically, the edge inference engine is optimized with model compression techniques such as weight pruning, quantization-aware training and knowledge distillation. This allows it to run on resource limited medical hardware with a loss in predictive performance. Clinically, the invention decreases the diagnostic turnaround time to virtually real-time (usually less than two seconds per scan) and enhances the efficiency of the workflow greatly in the high-volume screening programs. The automated triage feature helps radiologists by ranking the high-risk cases thus decreasing cognitive load and increasing diagnostic accuracy. The system architecture is also scalable in nature. Further IoT imaging nodes can be added without redesigning. Its modular structure allows updating to newer neural network models including transformer-based vision-based models without introducing changes in the communication infrastructure. Notably, the invention helps solve an age-old problem of the balancing of diagnostic sensitivity and specificity. The framework has a high discrimination performance over the traditional CAD systems because of its deep hierarchical characteristics, morphological validation, probabilistic calibration, explainable outputs. Overall, the suggested IoT-based edge -cloud deep learning platform creates the next-generation smart solution platform in the field of oncology diagnostics with the following features: • Localization of tumours in real-time. • Multi-class prediction of malignancy. • Fusion of morphological features. • Privacy-training distributed learning. • Decision outputs, which are clinically interpretable. • Continuous improvement which is adaptive. The invention is a holistic integration of IoT, edge AI, and medical imaging intelligence that can positively impact the early detection of breast cancer and precision oncology diagnostic.
Notices, Deadlines & Correspondence
Patent Information
Applicants
SR UNIVERSITY
Inventors
1. Dr. G. NAGA RAMA DEVI
2. Dr BALAJEE MARAM
Claims
1. A real-time breast tumour identification and classification system based on an IoT, including an imaging acquisition module, an edge computing unit, and a deep learning diagnostic core deployed on the edge computing unit and including a multi-scale convolutional neural network, configured to extract lesion-specific features, spatial attention mechanism, and a multi-classification engine configured to give probabilistic outputs identifying a tumour as benign, malignant, or intermediate-risk such that primary inference is done at the edge computing unit to enable the system to provide real-time diagnostic output.
2. The system in claim 1 whereby the preprocessing would entail noise filtering, contrast enhancement, intensity normalization, and automatic region-of-interest extraction.
3. According to claim 1, the system whose output is said to be in claim 1, the multi-scale convolutional neural network has parallel branches of convolution units, which are set against recognizing macro-structural abnormality and microcalcification patterns in breast tissue.
4. The system purported in claim 1, further including a morphological feature fusion module, which is programmed to incorporate deep learning features with structural descriptors such as: boundary irregularity index, compactness ratio, texture entropy and spiculation measures.
5. System as stated in claim 1, whereby the spatial attention system dynamically gives weights to spatial feature map to enhance the accuracy of localization of tumours, especially in dense breast tissues.
6. The system of claim 1, which also includes an explainable artificial intelligence component that is set up to produce lesion heatmaps, boundary overlays, and classification confidence visualization.
7. The system of claim 1, which also has a secure cloud synchronization layer which is programmed to accept encrypted model parameters or performance data to refine the adaptive model with but without sending raw patient imaging data.
8. This system according to claim 1, which involves the classification engine giving calibrated probability scores to aid in clinical risk stratification and in diagnostic decision-making.
9. The system as described in claim 1 which also includes a temporal monitoring module is also set to compare sequential patient scans to determine trends in tumour progression and growth.
10. A scheme to identify and characterize breast tumours in real-time that includes capturing a breast diagnostic image with an IoT-enabled imaging device; transmission of the diagnostic image to an edge computing unit; preprocessing the image to reveal diagnostic features; deriving hierarchical features with multi-scale convolutional neural network; applying a spatial attention mechanism in tumour-localization; fusing deep learning features with morphological descriptors; classifying the tumour as benign, malignant or intermediate-risk; and inferring probabilistic diagnostic results in real-time.
Specification
Description:The current invention is associated with an IoT-based real-time system of breast tumour identification and classification integrating smart imaging hardware, edge-based computation, and deep learning models in a hybrid edge and cloud system.
As shown in FIG. 1, the system will consist of four key functionalities, which are the following: an IoT-enabled imaging acquisition unit, an edge computing and preprocessing unit, a deep learning diagnostic core, and a secure cloud synchronization layer.
The IoT imaging acquisition is the module that comprises of smart mammography, ultrasound, or MRI equipment that is set up with secure communication interfaces. When an imaging device captures a breast image, the data is formatted in a standard medical imaging format and it is sent safely to a nearby edge computing unit. This local transmission reduces the latency, and this does not require immediate centralized cloud process.
The edge computing unit does real-time preprocessing to improve the accuracy of diagnostics. Such functions are noise reduction, contrast enhancement, intensity normalization, and automatic region-of-interest (ROI) extraction. The step of ROI extraction identifies the possible suspicious areas of the breast tissue so that the downstream neural network processing could be maximized.
The resulting processed image is then inputted into a multi-scale deep convolutional neural network (CNN) which comprises the heart diagnostic engine of the invention. The CNN architecture has parallel convolutional layers that are modelled to represent macro-architectural deformities as well as micro-scale calcifications. The analysis of tissue asymmetry, boundary irregularities, and textural variations can be effectively analysed using hierarchical feature aggregation to Analyze the system.
An attention mechanism of space is incorporated into intermediate levels of the neural system to highlight diagnostically meaningful areas and ignore background tissue noise. This enhances the sensitivity of detection especially in dense breast tissues whereby the visibility of tumours might be lower.
The system also includes a morphological feature fusion component, which integrates features obtained by deep learning with structural descriptors, which include compactness ratio, boundary irregularity, texture entropy, and spiculation measures. This mixed-field mechanism of validation minimizes false positives and improves the reliability of classification.
The classification subsystem is a multi-class probabilistic prediction scheme, which classifies lesions detected into benign, malignant or intermediate-risk lesions. A calibrated softmax layer produces confidence scores, which allow clinicians to determine diagnostic confidence.
In order to enhance interpretability, explainable AI outputs are produced by the system, such as lesion heatmaps and boundary overlays, that can visually show the regions that have the largest contribution to the classification decision.
Primary inference is run at the edge node, and this guarantees low-latency diagnostic output to be appropriate to real-time clinical processes. The system also provides optional encrypted synchronization of model parameters to a centralized cloud server enabling model refinement to be consistent continuously without the need to transmit the raw patient imaging data.
The scalability and modular nature of the invention facilitates its implementation in hospitals, diagnostic centres, and remote healthcare institutions and, therefore, improves the use of the
intelligent, privacy-conscious, and real-time diagnostic automation in early breast cancer detection. , Claims:1. A real-time breast tumour identification and classification system based on an IoT, including an imaging acquisition module, an edge computing unit, and a deep learning diagnostic core deployed on the edge computing unit and including a multi-scale convolutional neural network, configured to extract lesion-specific features, spatial attention mechanism, and a multi-classification engine configured to give probabilistic outputs identifying a tumour as benign, malignant, or intermediate-risk such that primary inference is done at the edge computing unit to enable the system to provide real-time diagnostic output.
2. The system in claim 1 whereby the preprocessing would entail noise filtering, contrast enhancement, intensity normalization, and automatic region-of-interest extraction.
3. According to claim 1, the system whose output is said to be in claim 1, the multi-scale convolutional neural network has parallel branches of convolution units, which are set against recognizing macro-structural abnormality and microcalcification patterns in breast tissue.
4. The system purported in claim 1, further including a morphological feature fusion module, which is programmed to incorporate deep learning features with structural descriptors such as: boundary irregularity index, compactness ratio, texture entropy and spiculation measures.
5. System as stated in claim 1, whereby the spatial attention system dynamically gives weights to spatial feature map to enhance the accuracy of localization of tumours, especially in dense breast tissues.
6. The system of claim 1, which also includes an explainable artificial intelligence component that is set up to produce lesion heatmaps, boundary overlays, and classification confidence visualization.
7. The system of claim 1, which also has a secure cloud synchronization layer which is programmed to accept encrypted model parameters or performance data to refine the adaptive model with but without sending raw patient imaging data.
8. This system according to claim 1, which involves the classification engine giving calibrated probability scores to aid in clinical risk stratification and in diagnostic decision-making.
9. The system as described in claim 1 which also includes a temporal monitoring module is also set to compare sequential patient scans to determine trends in tumour progression and growth.
10. A scheme to identify and characterize breast tumours in real-time that includes capturing a breast diagnostic image with an IoT-enabled imaging device; transmission of the diagnostic image to an edge computing unit; preprocessing the image to reveal diagnostic features; deriving hierarchical features with multi-scale convolutional neural
network; applying a spatial attention mechanism in tumour-localization; fusing deep learning features with morphological descriptors; classifying the tumour as benign, malignant or intermediate-risk; and inferring probabilistic diagnostic results in real-time.
Documents
Application Documents
| # | Name | Date |
|---|---|---|
| 1 | 202641027222-FORM-9 [09-03-2026(online)].pdf | 2026-03-09 |
| 2 | 202641027222-FORM 1 [09-03-2026(online)].pdf | 2026-03-09 |
| 3 | 202641027222-FIGURE OF ABSTRACT [09-03-2026(online)].pdf | 2026-03-09 |
| 4 | 202641027222-DRAWINGS [09-03-2026(online)].pdf | 2026-03-09 |
| 5 | 202641027222-COMPLETE SPECIFICATION [09-03-2026(online)].pdf | 2026-03-09 |