Artificial intelligence platform for multi-modal radiological image analysis, cardiac MRI quantification, pneumonia classification, tumor segmentation and grading

Abstract

The invention discloses an artificial intelligence platform for comprehensive radiological image analysis across CT, MRI, and X-ray modalities. The system automatically detects and quantifies cardiac, pulmonary, and oncological abnormalities through multiple AI modules including heart localization, cardiac MRI quantification, pneumonia classification, and tumor segmentation / grading. The platform provides automated tumor detection, 3D volumetric segmentation, and malignancy grading and treatment planning. The system integrates hybrid CNN-Transformer networks, attention-based 3D segmentation, and multi-modal feature fusion, offering precise quantitative metrics for tumor size, volume, shape, and malignancy probability. It also a cardiac MRI analysis module measures atrial and ventricular volumes and detects atrial fibrillation or mitral valve stenosis. It calculates the cardiothoracic ratio and detects heart displacement. DICOM / PACS integration and federated learning enable clinical deployment and cross-institutional adaptability while preserving patient data privacy and assists clinicians in accurate and rapid diagnosis while minimizing human error. This first-in-world platform facilitates early detection, staging, grading, and treatment planning with high reproducibility and explainable outputs.

Description

[0001] Artificial Intelligence Platform for Multi-Modal Radiological Image Analysis, Cardiac MRI Quantification, Pneumonia Classification, Tumor Segmentation and Grading

[0002] TECHNICAL FIELD

[0003] The present disclosure generally relates to the field of digital health, artificial intelligence, and radiological image analysis in medical screening. More specifically, it pertains to an Al-driven platform designed for the tumor segmentation and grading; cardiac MRI quantification, and pneumonia classification, employing deep learning-based algorithms for automated detection, localization, and quantification. By leveraging Al-powered digital analysis, the disclosed invention enhances the diagnosis and detection, improves assessment accuracy, and optimizes healthcare resources. The innovative software framework employs a three-tier architecture, comprising a presentation layer for user interaction, a business logic layer for data processing, and a data access layer for managing system data. This holistic approach addresses current limitations in screening and diagnosis programs, improves accessibility to predictive diagnostics, and healthcare management.

[0004] BACKGROUND

[0005] Existing computer-aided diagnostic systems are often limited to specific pathologies or single imaging modalities, lacking comprehensive integration to diagnosis different conditions and realtime quantitative capabilities. Conventional radiological image interpretation remains heavily dependent on the expertise of radiologists. Despite advances in imaging hardware, diagnostic accuracy and consistency vary considerably among practitioners. Human visual assessment is inherently limited by fatigue, subjective interpretation, and workload pressures. These constraints lead to delays in diagnosis, inter-observer variability, and missed findings.

[0006] In cardiac imaging, accurate assessment of atrial and ventricular volumes using MRI is critical for identifying functional disorders such as atrial fibrillation, dilated cardiomyopathy, and valvular stenosis. Manual contouring of cardiac chambers remains a labor-intensive and time-consuming task requiring specialized training. Current Al approaches in this area either depend on pre-defined segmentation templates or limited 2D slice analysis, which fail to generalize to full 3D volumetric datasets or account for cardiac motion variability and tumor grading.

[0007] Recent advances in computer-aided diagnostic (CAD) systems and Al have led to automated tumor detection and segmentation algorithms. These systems often use convolutional neural networks (CNNs) for 2D slice analysis or simple 3D volume reconstruction. However, existing Al systems are generally limited in several key ways including single-modality focus, lacking flexibility to integrate multiple imaging modalities such as PET / CT or multiphase MRI, limited volumetric precision and staging estimation, static models, wherein Al models require retraining for new data sources, lacking federated or adaptive learning frameworks.

[0008] Consequently, there exists a significant unmet need for a unified, multi-modal Al platform that provides comprehensive tumor segmentation and grading, robust volumetric quantification, and interpretability for clinical decision support. Accurate determination of cardiac size and position, measurement of atrial and ventricular volumes in cardiac MRI, identification of pneumonia in chest X-rays, and tumors segmentation in CT scans are all critical tasks in diagnostic radiology. The absence of an integrated, intelligent system capable of performing all these analyses across modalities leads to inefficiencies, diagnostic delays, and potential oversight of life-threatening abnormalities.

[0009] SUMMARY

[0010] According to this invention, there is provided an Al-driven platform designed to multi-modal radiological image analysis, offering automated detection, localization, segmentation, and diagnostic assistance for cardiopulmonary and oncological diseases. The invention includes several distinct but interoperable modules: a preprocessing module for noise and contrast normalization; a cardiac localization and bounding-box generation module using transformerbased vision models; a cardiac MRI quantification module leveraging 3D convolutional and recurrent layers to compute atrial / ventricular volumes dynamically; a pneumonia classification module trained on multimodal feature embeddings; and a lung tumor segmentation module utilizing hybrid CNN-transformer architectures for volumetric delineation. Each module is orchestrated within a federated learning framework that enables distributed model adaptation across clinical institutions while preserving patient data privacy.

[0011] The system is implemented on GPU-accelerated servers or cloud infrastructure. Radiological images are ingested via DICOM interfaces or PACS retrieval. After preprocessing, the Al pipeline performs multi-modal feature extraction, tumor detection, segmentation, and grading. Output masks and grading reports are rendered in a clinician-friendly viewer with overlays and quantitative metrics. The system continuously adapts through federated learning, allowing deployment across multiple institutions while maintaining data privacy.

[0012] According to this invention, there is provided a system that leverages advanced machine learning techniques and CNN-transformer architectures to employs deep convolutional or transformerbased neural networks to automatically detect the position and size of the heart and to generate an adaptive bounding box with precise margins.

[0013] It calculates metrics such as the cardiothoracic ratio and heart displacement index, supporting detection of cardiomegaly and related abnormalities.

[0014] It automatically measures end-diastolic and end-systolic volumes, ejection fraction, and atrial size, and identifies abnormalities including atrial fibrillation and mitral valve stenosis.

[0015] According to this invention, there is provided a method and applies 3D segmentation algorithms on CT / MRI images to detect and delineate tumors to measures tumor volume, spatial coordinates, and generates data relevant to tumor staging (TNM classification).

[0016] According to this invention, there is provided seamless interoperability with DICOM and PACS systems, allowing users to upload or retrieve images directly from medical imaging archives.

[0017] The system supports federated learning, enabling distributed model training without centralizing patient data, ensuring both privacy and continuous model improvement.

[0018] DETAILED DESCRIPTION OF THE INVENTION

[0019] The present invention discloses a "Smart PACS" platform and the world’s first integrated system that overcomes the limitations of the prior art by providing a holistic system for intelligent medical image management and analysis. The core of the invention is the integration of multiple, specialized Al analysis modules directly into a DICOMweb-compliant PACS framework. The key components and features of the invention include:

[0020] 1. A Centralized Web-Based PACS Platform: A server-based system that manages and serves medical images in compliance with the DICOM standard.

[0021] 2. A Plurality of Integrated Al Diagnostic Models: The platform natively integrates several distinct, pre-trained Al models, each designed for a specific diagnostic task, including but not limited to: o A Tumor Segmentation Model for identifying and delineating tumor boundaries in MRI / CT scans, enabling accurate volume measurement and tumor grading module that integrates radiomic, morphological, and PET-derived metabolic features via CNN, Vision Transformer (ViT), Radiomics + ML (SVM, XGBoost), Hybrid CNN-ML models to identify tumour masses, segmentation, volumetric quantification, and grading to guide surgical or chemotherapeutic interventions. o A Cardiac Detection Model for automatically localizing the heart and drawing a bounding box around it in radiographs to aid in detecting cardiomegaly and related pulmonary conditions. o A Left Atrium Segmentation Model for precisely segmenting the left atrium in Cardiac MRI scans to quantify volume and assist in diagnosing conditions like atrial fibrillation. o A Pneumonia Classification Model for automatically classifying Chest X-Ray images to identify the presence or absence of pneumonia based on opacities and other visual indicators.

[0022] 3. An Integrated DICOM Viewer with Analysis Toolbox: A web-based viewer that is natively integrated with the platform, providing tools for image manipulation (e.g., windowing, zooming, panning) and measurement, and serving as the interface for launching and displaying results from the Al models.

[0023] 4. Pull DICOMweb Protocol Support: The platform implements the DICOMweb standard, including: o QIDO-RS (Query based on ID for DICOM Objects by RESTful Services) for efficient, metadata-based searching and filtering of studies. o WADO-RS (Web Access to DICOM Persistent Obj ects by RESTful Services) for retrieving images, metadata, and bulk data in a modern, web-friendly manner.

[0024] 5. Unified Workflow: The invention enables a unified diagnostic workflow wherein a user can query studies via QIDO-RS, retrieve them via WADO-RS, view them in the integrated viewer, and execute one or more Al analyses with a single click, with results overlaid directly on the original images within the same interface.

[0025] System Overview:

[0026] The platform comprises a unified software and computational framework configured to receive medical imaging data (including, but not limited to, MRI, CT, PET, or hybrid modalities such as PET / CT or PET / MRI), process the data through modular Al pipelines, and output clinically relevant diagnostic metrics.

[0027] Each module operates autonomously or in coordination with other system components through a centralized controller and standardized data exchange interface. The modules may be deployed on local hospital servers, cloud-based processing environments, or edge computing devices within imaging scanners.

[0028] The Al algorithms are trained on annotated medical datasets using supervised, semi-supervised, or transfer learning approaches to ensure robust generalization across imaging protocols, patient demographics, and hardware vendors.

[0029] Tumor Grading Module:

[0030] The tumor grading module integrates radiomic, morphological, and PET-derived metabolic features within a unified analytical framework to identify tumor masses, perform segmentation, quantify volumetric attributes, and generate malignancy grading outputs to guide surgical or chemotherapeutic interventions.

[0031] Image Preprocessing and Registration:

[0032] The module receives multi-parametric imaging inputs — such as T1-, T2-, FLAIR-weighted MRI, CT, and PET images — and performs automatic preprocessing steps including:

[0033] Intensity normalization,

[0034] Noise suppression and artifact correction,

[0035] Spatial alignment (registration) across modalities, and

[0036] Region-of-interest (ROI) extraction.

[0037] This ensures that anatomical and metabolic features are co-registered for subsequent feature extraction and model inference.

[0038] Tumor Segmentation and Volumetric Quantification:

[0039] Segmentation is performed via Convolutional Neural Network (CNN) and Vision Transformer (ViT) architectures. In some embodiments, a 3D U-Net or nnU-Net configuration is employed to delineate tumor boundaries voxel-by- voxel, while the Vision Transformer captures long-range contextual relationships within the volumetric data.

[0040] Post-processing operations (morphological filtering, connected component analysis) are applied to refine the segmentation masks. Tumor volume is calculated based on voxel count multiplied by voxel dimensions, yielding precise volumetric measurements. Leature Extraction:

[0041] The module extracts multiple categories of features:

[0042] Radiomic features: texture, intensity, and wavelet- based descriptors derived from segmented regions;

[0043] Morphological features: shape descriptors such as sphericity, compactness, irregularity, and surface-area-to-volume ratio; and

[0044] Metabolic features: PET-derived standardized uptake values (SUVmean, SUVmax, total lesion glycolysis).

[0045] These feature sets are normalized and concatenated into a multi-dimensional feature vector for integrated analysis.

[0046] Multi-model integration for tumor grading

[0047] To perform tumor grading and malignancy scoring, the module implements a hybrid Al approach combining:

[0048] Deep learning models such as CNN and Vision Transformer (ViT) for spatial feature extraction;

[0049] Machine learning classifiers, including Support Vector Machine (SVM) and XGBoost, trained on radiomic and morphological feature vectors; and

[0050] Hybrid CNN-ML pipelines, in which CNN-derived embeddings are fed into ML classifiers to enhance explainability and reduce overfitting.

[0051] The resulting classifier outputs a malignancy probability score and an assigned tumor grade (e.g., WHO Grade I-IV or equivalent), which are displayed in an interpretable report for clinician review.

[0052] Clinical Integration:

[0053] The tumor grading module interfaces with hospital information systems (HIS) and picture archiving and communication systems (PACS) to export segmentation masks, volumetric data, and grading reports. The output guides treatment planning by informing:

[0054] Surgical margin estimation,

[0055] Radiotherapy dose mapping, and

[0056] Chemotherapy regimen selection.

[0057] Cardiac Localization Module The cardiac localization module performs automated localization and morphometric analysis of cardiac anatomy to calculate the cardiothoracic ratio (CTR) and detect heart displacement using Al model combinations. The module operates on 2D or 3D medical images, such as chest X- rays, CT, or MRI datasets.

[0058] Image Segmentation and Detection:

[0059] The module employs deep neural architectures, including nnU-Net, DeepLabV3+, and Y0L0v8, to segment or detect cardiac and thoracic boundaries:

[0060] The nnU-Net architecture provides pixel-wise segmentation of the heart and thoracic cage;

[0061] DeepLabV3+ enhances boundary precision using dilated convolutions and attention mechanisms;

[0062] Y0L0v8 enables real-time detection of cardiac bounding boxes and key anatomical landmarks (e.g., cardiac apex, midline, diaphragm).

[0063] The output segmentation masks or bounding boxes are used to compute anatomical distances relevant to CTR and heart displacement.

[0064] Cardiothoracic Ratio Computation:

[0065] Using the segmentation outputs, the module automatically determines:

[0066] Maximum heart width (transverse cardiac diameter),

[0067] Maximum thoracic width (distance between inner rib cage boundaries), and computes the Cardiothoracic Ratio (CTR) as: CTR=Thoracic Width / Cardiac Width

[0068] The computed ratio is compared against clinically accepted thresholds (e.g., CTR > 0.5 indicative of cardiomegaly).

[0069] Heart Displacement Detection:

[0070] Heart displacement or mediastinal shift is detected using an Al classification subsystem that combines:

[0071] Feature-based ML models such as XGBoost, utilizing geometric and positional features (e.g., heart center offset, lung asymmetry); and

[0072] Deep CNN classifiers such as ResNet, trained to recognize displacement patterns directly from imaging data. These model combinations allow robust differentiation between normal and abnormal cardiac positioning, even under varying patient orientations or imaging conditions.

[0073] Diagnostic Output:

[0074] The cardiac localization module outputs:

[0075] The CTR value (numeric),

[0076] A classification label indicating “normal,” “cardiomegaly,” or “displacement detected,”

[0077] Annotated visualization overlays showing heart and thorax boundaries.

[0078] These outputs assist cardiologists and radiologists in identifying cardiac enlargement or mediastinal shift and in monitoring disease progression.

[0079] The invention provides several technical and clinical advantages, including: o Fully automated, multi-organ diagnostic capability integrating oncology and cardiology analyses within one platform. o High diagnostic precision achieved through hybrid deep learning and radiomics-based ML architectures. o Quantitative outputs (tumor volume, CTR, displacement vectors) to support objective and reproducible clinical decision-making. o Scalable deployment across imaging modalities and institutions via modular design. o Explainability and regulatory compliance, enhancing physician trust and traceability in AI- assisted diagnostics. o Load and show instance of data:

[0080] Cardiac Detection import os import numpy as np import torch import torchvision import cv2 import pydicom from PIL import Image, ImageDraw import matplotlib.pyplot as pit class CardiacDetectionModel(torch.nn.Module): def init (self): superQ. init () selfmodel = torchvision.models.resnetl8(pretrained=False) selfmodel.convl = torch. nn.Conv2d(l, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) selfmodel. fc = torch. nn.Linear(in_features=512, out_features=4) def forward(self, data): return self.model(data) def load_model(model_path):

[0081] > Load the trained cardiac detection model device = torch. device("cuda" if torch. cuda.is_available() else "cpu") model = CardiacDetectionModelQ model. load_state_dict(torch. load(model_path, map_location=device)) model. eval() return model, device def preprocess_dicom(dicom_path, target_size=224):

[0082] > Preprocess DICOM image for model input

[0083] # Read DICOM file dem = pydicom.dcmread(dicom_path) dem array = dcm.pixel array original height, original width = dem array, shape

[0084] # Resize and normalize for model input dcm_array_resized = cv2.resize(dcm_array, (target_size, target_size)) dcm_array_resized = (dcm_array_resized / 255).astype(np.float32) # Normalize with dataset statistics dcm_array_resized = (dcm_array_resized - 0.494) / 0.252

[0085] # Convert to tensor tensor = torch. tensor(dcm_array_resized).unsqueeze(0).unsqueeze(0) return tensor, dem array, original height, original width def detect_cardiac(model, device, input tensor):

[0086] > Detect cardiac region using the model with torch. no gradQ: input tensor = input tensor.to(device) prediction = model(input tensor) return prediction. cpu().numpy()[0] def scale_bbox_coordinates(bbox_coords, original size, target_size=224):

[0087] > Scale bounding box coordinates from model output to original image size xO, yO, xl, yl = bbox_coords orig_h, orig_w = original_size

[0088] # Calculate scaling factors scale_x = orig_w / target_size scale_y = orig_h / target_size

[0089] # Scale coordinates x0_scaled = int(x0 * scale_x) y0_scaled = int(y0 * scale_y) xl scaled = int(xl * scale x) yl _scaled = int(yl * scale_y) return [x0_scaled, y0_scaled, xl_scaled, yl_scaled] def resize_image_with_bbox(image_array, bbox coords, output_size=500):

[0090] . Resize image to desired output size while maintaining bounding box coordinates orig_h, orig_w = image_array. shape

[0091] # Calculate scaling factors scale = min(output_size / orig_w, output_size / orig h) new_w = int(orig_w * scale) new_h = int(orig_h * scale)

[0092] # Resize image if isinstance(image_array, np.ndarray): resized_image = cv2.resize(image_array, (new_w, new_h)) else: resized_image = image_array.resize((new_w, new_h), Image.Resampling.LANCZOS)

[0093] # Scale bounding box coordinates xO, yO, xl, yl = bbox_coords x0_scaled = int(x0 * scale) y0_scaled = int(y0 * scale) xl scaled = int(xl * scale) yl _scaled = int(yl * scale) return resized_image, [x0_scaled, y0_scaled, xl_scaled, yl_scaled] def draw_bounding_box(image_array, bbox coords): > Draw bounding box on the image

[0094] # Convert to PIL Image if it's a numpy array if isinstance(image_array, np.ndarray): if image_array.dtype != np.uint8:

[0095] # Normalize to 0-255 range image_array = ((image_array - image_array.min()) /

[0096] (image_array.max() - image_array.min()) * 255).astype(np.uint8) pil image = Image, fromarray(image array) else: pil_image = image_array

[0097] # Draw bounding box draw = ImageDraw.Draw(pil image) xO, yO, xl, yl = bbox_coords draw.rectangle([xO, yO, xl, yl], outline- ' orange", width=3) return pil image def process_dicom_image(model, device, dicom_path, output_path, output_size=500):

[0098] > Process DICOM image and save result with bounding box at desired size

[0099] # Preprocess DICOM input tensor, original array, orig h, orig w = preprocess_dicom(dicom_path)

[0100] # Get prediction (coordinates are for 224x224 image) bbox_coords_224 = detect_cardiac(model, device, input tensor) bbox_coords_224 = [int(coord) for coord in bbox_coords_224]

[0101] # Scale coordinates to original image size bbox coords original = scale_bbox_coordinates(bbox_coords_224, (orig h, orig w))

[0102] # Resize image and bounding box to desired output size resized image, bbox coords resized = resize_image_with_bbox( original array, bbox coords original, output size

[0103] )

[0104] # Draw bounding box on resized image result image = draw_bounding_box(resized_image, bbox coords resized)

[0105] # Save result result image. save(output_path) return output_path, bbox coords resized

[0106] Atrium Segmentation import torch import pytorch lightning as pl class DoubleConv(torch.nn. Module): def init (self, in channels , out channels): super (). init () selfstep = torch.nn.Sequential(torch.nn.Conv2d(in_channels , out channels , 3 , padding =1), torch. nn.ReLUQ, torch. nn.Conv2d(out_channels , out channels , 3 , padding =1), torch. nn.ReLUQ) def forward(self , X): return self.step(X) class UNet(torch.nn.Module): def init (self): super (). init () selflayerl = DoubleConv(l,64) self, lay er2 = DoubleConv(64,128) self.layer3 = DoubleConv( 128,256) self.layer4 = DoubleConv(256,512) self. Iayer5 = DoubleConv(512+256,256) self.layer6 = DoubleConv(256+128,128) self, layer? = DoubleConv( 128+64,64) self. Iayer8 = torch. nn.Conv2d(64, 1, 1) selfmaxpool = torch.nn.MaxPool2d(2) def forward(self , x): xl = self, lay er l(x) xlm = self.maxpool(xl) x2 = self.layer2(xlm) x2m = self.maxpool(x2) x3 = self, lay er3(x2m) x3m = self.maxpool(x3) x4 = self, lay er4(x3m) x5 = torch.nn.Upsample(scale_factor = 2 , mode- ' bilinear" )(x4) x5 = torch. cat([x5,x3] , dim = 1) x5 = self. Iayer5 (x5) x6 = torch.nn.Upsample(scale_factor = 2 , mode="bilinear")(x5) x6 = torch. cat([x6,x2] , dim = 1) x6 = self, lay er6(x6) x7 = torch.nn.Upsample(scale_factor = 2 , mode- ' bilinear" )(x6) x7 = torch. cat([x7,xl] , dim = 1) x7 = self, lay er7(x7) ret = self, lay er8(x7) return ret class DiceLoss(torch.nn. Module): def init (self): super (). init () def forward(self , pred , mask): pred = torch, flatten(pred) mask = torch, flatten(mask) counter = (pred * mask).sum() denum = pred.sum() + mask.sum() + le-8 dice = (2* counter) / denum return 1-dice class AtriumSegmentation(pl.LightningModule): def init (self): super (AtriumS egmentati on, self) . init () selftraining step outputs = [] selfvalidation step outputs = [] self, model = UNet() self. optimizer = torch. optim.Adam(self.model.parameters(), Ir = le-4) self loss fn = DiceLoss() def forward(self , data): return torch. sigmoid(self. model(data)) def training_step(self , batch , batch idx): mri, mask = batch mask = mask.float() pred = self(mri) loss = selfloss_fn(pred , mask) selftraining step outputs.append(loss) self. log(" Training Dice", loss) if batch idx % 50 ==0 : self.log_images(mri.cpu() , pred.cpu() , mask.cpu() , "Train" ) return loss def validation_step(self , batch , batch idx): mri, mask = batch mask = mask.float() pred = self(mri) loss = self.loss_fn(pred , mask) self.validation step outputs.append(loss) self.log("Val Dice", loss) if batch_idx % 2 =0 : self log_images(mri.cpu() , pred.cpu() , mask.cpu() , "Vai" ) return loss def log_images(self, mri , pred , mask , name): pred = pred > 0.5 fig , axis = plt.subplots(l,2) axis[0].imshow(mri[0][0] , cmap- 'bone") mask_ = np.ma.masked_where(mask[0][0]==0 , mask[0][0]) axis[0].imshow(mask_ , alpha = 0.6) axis[l].imshow(mri[0][0] , cmap- 'bone") mask_ = np.ma.masked_where(pred[0][0]==0 , pred[0][0]) axis[l].imshow(mask_ , alpha = 0.6) self. logger. experiment. add_figure(name , fig , self.global_step) def configure optimizers(self): return [self, optimizer]

Claims

ClaimsWhat is claimed is:

1. A comprehensive artificial intelligence platform for multi-modal radiological image analysis comprising: o a preprocessing module configured for image normalization; o a neural network-based localization module configured to detect the heart and generate a bounding box; o a preprocessing module for image normalization and modality-specific feature extraction; o a tumor detection module employing a hybrid CNN-Transf ormer architecture; o a 3D volumetric segmentation module with attention-based skip connections; o a tumor grading module predicting malignancy probability scores; o a quantitative volumetric changes for evaluating therapy response; o a segmentation module configured to delineate cardiac, pulmonary, and oncological structures; and o an integration module configured to interface with DICOM and PACS systems.

2. The platform of claim 1, wherein the tumor grading module integrates radiomic, morphological, and PET-derived metabolic features via CNN, Vision Transformer (ViT), Radiomics + ML (SVM, XGBoost), Hybrid CNN-ML models to identify tumour masses, segmentation, volumetric quantification, and grading to guide surgical or chemotherapeutic interventions.

3. The platform of claim 1, wherein the output includes attention maps, segmentation masks, tumor volume, shape metrics, and malignancy probability scores to increase diagnostic capacity, segmentation, tumor volume assessment, malignancy grading for staging and treatment planning.

4. The platform of claim 1, wherein a cardiac MRI analysis module via cNeural Networks (CNNs) and U-Net variants for segmentation (2D, 2.5D, or 3D) measures atrial and ventricular volumes and detects atrial fibrillation or mitral valve stenosis.

5. The platform of claim 1, wherein the cardiac localization module via nnU- Net / DeepLabV3+ / YOLOv8 and XGBoost / ResNet algorithm combinations calculates the cardiothoracic ratio and detects heart displacement.

6. The platform of claim 1, wherein the federated learning framework enables distributed training without sharing raw patient data.

7. The platform of claim 1, wherein the system is capable of analyzing CT, MRI, and PET / CT scans interchangeably to provide clinicians with reproducible, interpretable results, reducing diagnostic errors and variability8. The platform of claim 1, wherein a pneumonia classification module detects and classifies radiographic signs of pneumonia in chest X-rays.

9. The platform of claim 1, wherein a lung tumor segmentation module detects and quantifies tumor size and volume in CT images for staging analysis.

10. The platform of claim 1, wherein the system provides visualization overlays including bounding boxes, heatmaps, and segmentation contours on a DICOM viewer.

11. The platform of claim 1, wherein the system employs deep learning models selected from convolutional neural networks, transformers, or hybrid architectures.

12. The platform of claim 1, wherein model training is performed using federated learning for continuous adaptation while preserving patient privacy.

Images