Description:FIELD OF THE INVENTION:
The present invention relates to a system for obtaining clinical inference from thermograms. More particularly, the present invention relates to an interpretable computer assisted decision support system that facilitates clinical inference of breast abnormalities by accompanying a prediction with visual identification of the region of interest (ROI) on the breast thermography images.

BACKGROUND OF THE INVENTION:
Breast cancer is one of the most lethal diseases in women, and its prevalence is increasing at an alarming rate. It is a type of cancer that develops in the breast cells and progresses in stages. The mortality rate of breast cancer in the remote regions of the country is notably higher due to lack of scalable and affordable breast screening systems for clinical prediction of abnormality, exacerbated by inadequately equipped facilities for timely diagnosis and treatment.

The popular imaging modalities such as Radio Wave Imaging, X-ray Imaging, Ultrasound Imaging, Gamma Imaging, Radio-nuclide Imaging, and Functional Imaging cannot be used for breast cancer screening in the rural areas, because they require expert health personnel to operate who are willing to serve only in the well-equipped tertiary hospitals. Furthermore, these modalities are expensive and painful involving human exposure to radiation and most of the devices involved are not portable. However, Infrared Imaging modality gets heightened attention for early breast screening as this is a portable, low cost, non-invasive, pain-free, and radiation-free technique. Infrared (IR) imaging uses electromagnetic radiation (EMR) with wavelengths longer than those of visible light and shorter than radio waves, to produce an image of a region of interest. In this technique, the infrared radiation emitted by a human body is captured by using a thermal camera.

It has been reported that the growth rate of a tumor has a correlation with temperature. This correlation is utilized in Thermography for detecting the possible location of tumors in the body. Thermography is a technique for detecting and measuring variations in the heat emitted by various regions of the body and transforming them into visible signals that can be recorded photographically (as for diagnosing abnormal or diseased underlying conditions). The heat patterns and blood flow of the body tissues are captured by an infrared camera. It exploits thermal symmetry of the human body, thereby detecting temperature variability in the cancerous region of the body. Evidence of changes in symmetrical thermal patterns aids in the detection of abnormalities in a healthy patient by identifying warmer cancerous regions, even during the angiogenesis phase.

Computer-assisted diagnosis of cancer using thermography technique offers a promising solution to reduce the mortality rates in rural India. This becomes even more important since there is a lack of sufficient expertise and infrastructure in rural regions. However, development of a computer assisted expert system for breast abnormality detection using thermograms is challenging in rural areas, since sufficient number of breasts thermograms in Indian scenario are not openly available to develop an expert Computer aided Design (CAD) system in these areas. Besides, the resulting thermograms are primarily littered with noisy intrusions, emanated by atmospheric temperature changes resulting in low Signal to Noise Ratio (SNR), which produces false positives and false negatives in the classification of the abnormalities in breast tissues.

Most imaging techniques, including thermography, perform accurate and precise predictions of a region based on the extracted features from the images. The extraction of relevant features from precise breast region of interest (ROI) results in fewer false positives and false negatives in thermogram analysis. However, it is noted that failure to identify the best intrinsic features results in the loss of relevant information from the thermal images, which leads to poor classification performance.

Advances in the field of artificial intelligence (AI) address each of these challenges and have greatly aided in the development of an effective CAD system for detecting abnormalities in thermography images. Convolutional Neural Network (CNN) has been shown to be the most conformable approach for recognition of abnormality in thermography images. In particular, the Convolution Neural Network (CNN) based Deep Learning techniques have proven to be promisingly effective in identifying the abnormal regions in thermography images. Deep Learning methods have gained prominence in pattern recognition, wherein relevant features and complex patterns are automatically extracted for efficient analysis of images.

However, an important challenge with CNN is its need for massive amounts of data to produce satisfactory results in the biomedical image analysis. Despite their high sensitivity and predictive power, their clinical translation is hampered due to a lack of interpretable insights about the prediction. Transfer learning technique provides a propitious solution to overcome this drawback, and assures better performance of CNN based methods, even with relatively small sized datasets.

In transfer learning, a network trained to one domain is finetuned and deployed to a different domain application. CNN models trained on a large-scale ImageNet dataset are widely adopted in the classification of breast thermography images, using transfer learning. They have been consistent in the successful detection of breast abnormality by analyzing thermography images due to their simplistic in-depth structure. However, despite the promising performance of the CNN models in thermography image analysis, there is lack of clarity regarding which segment of the image contributes to a particular decision that is arrived at; thus, raising a question mark on the confidence level of the decision.

The introduction of data visualization powered by Artificial Intelligence (AI) drives the diagnostic interpretability with CNN prediction. It reveals underlying causes for decisions and promotes the reliability of the CNN model for critical applications such as abnormality detection from breast regions in the thermograms.

Despite the high mortality rate of breast cancer, there has been very limited research into integrating screening techniques with machine learning approaches to provide sufficient expertise and real time communication to remote regions for in time diagnosis. Therefore, there is a need for an affordable and portable breast screening system for thermography images, using infrared imaging to facilitate breast health monitoring in the developing part of the country. This computer-assisted framework using thermography breast images would greatly benefit healthcare professionals for a precise and timely diagnosis of abnormality in the thermograms.

OBJECT OF THE INVENTION:
In order to obviate the drawbacks of the existing state of the art, the present invention discloses a system for obtaining clinical inference from thermography images by accompanying a prediction with visual identification of thermography to enable recognition of abnormality from the said breast thermogram.

The main object of the present invention is to provide a system for obtaining clinical inference from thermograms by an interpretable computer assisted expert model. More specifically, the object is to provide an ABN-DCN model which integrates a DarkCovidNet (DCN) feature extractor with an Attention Branch Network (ABN) for binary classification of thermograph images, by accompanying a prediction with visual identification of the region of interest (ROI) on the said thermography images.

Another object of the invention is to formulate a breast thermogram dataset such as the Amrita Breast Thermogram (ABT) dataset for image acquisition using a thermal camera and associated software.

Another object of the invention is to utilize Attention Branch Network (ABN) that introduces an attention branch using response based visual explanation, to emphasize diagnostic interpretability by identifying the region of interest (ROI) of thermographic images, thus enabling binary classification of breast thermography images.

Another object of the invention is to generate a heatmap to identify the region of interest (ROI) in the thermography images by using the attention branch that applies a convolution layer of DarkNet19 and Global Average Pooling (GAP) to model the context of the visual features of the thermogram.

Another object of the invention is to employ DarkNet19, a well-known pre-trained variant of DarkCovidNet (DCN), as a feature extractor for ABN to capture and learn the relevant features from histopathology images which can have a substantial impact on the interpretability of classification.

Another object of the invention is to mitigate dataset imbalance in thermography images through the implementation of sample weighting techniques, thereby alleviating model bias favoring the dominant class. In the Amrita Breast Thermogram (ABT) dataset, there are more images in the normal class.

Another object of the invention is to classify the thermography images into normal and abnormal classes using a perception branch (PB).in order to output the final probability for each class.

SUMMARY OF THE INVENTION:
The present invention discloses a system for obtaining clinical inference from thermograms. The invention discloses an interpretable computer assisted expert decision support system to recognize abnormality from breast thermogram. The system utilizes a new dataset Amrita Breast Thermogram (ABT) dataset for the development of CAD system recognizing abnormality in the breast thermogram. An efficient thermal camera-based system is set up for the acquisition of breast images in Indian scenario collaborating with super specialty hospital. The system introduces a new potential approach in diagnostic interpretability by using Attention Branch Network (ABN) that introduces an attention branch, to enable response based visual explanation, for identifying the region of thermogram where the Convoluted Neural Network (CNN) model focusses. This facilitates the binary classification of breast thermography images.

A modified DarkNet19 model, which is a pre-trained variant of DarkCovidNet (DCN), has been used as a feature extractor for ABN to capture the best intrinsic features from histopathology images which can have a substantial impact on the interpretability of classification. The extracted features from thermography images are used to prepare feature maps (FM). The DarkNet19 has been adopted as the base model, because of its lesser depth and complexity. The attention branch of ABN is integrated with the said DCN feature extractor to generate attention maps (AM) using attention mechanisms.

Finally, the attention maps (AM) and feature maps (FM) are fed to the perception branch to output the probability of each class, facilitating the classification of thermography images. The adaptability of the proposed CNN model is further strengthened by data augmentation on the training data. This enables synthetical expansion of the data pool, thus inherently helping reduce over-fitting which improves the capability of the model to generalize the training set.

Additionally, the system addresses the bias of the present in the model by incorporating a technique known as sample weighting, specifically focusing on the minority class within the dataset. This method involves assigning higher importance to instances from the underrepresented class, effectively allowing the model to learn and generalize better from these examples, ultimately leading to a more balanced and unbiased model performance.

BRIEF DESCRIPTION OF THE DRAWINGS:
Fig. 1 depicts the Overall flow diagram of the interpretable computer assisted expert model.
Fig. 2 depicts the Thermal camera-based image acquisition set-up.

Fig. 3 depicts image samples in the Enhanced Amrita Breast Thermogram (ABT)
dataset

Fig. 4 depicts Architecture of the decision-support ABN-DCN model.

Fig. 5 depicts the attention map generated by the ABN-DCN model.

DETAILED DESCRIPTION OF THE INVENTION:
The present invention discloses a system (S) for obtaining clinical inference from thermograms. The invention discloses an interpretable computer assisted expert system for the recognition of abnormality from breast thermograms (BT). The interpretable computer assisted decision support system of the invention facilitates clinical inference of thermography images by accompanying a prediction with visual identification of the region of interest (ROI) on the thermography images.

The system, as disclosed, utilizes an ABN-DCN model. The model integrates an attention branch (AB) of Attention Branch Network (ABN) with a DarkCovidNet (DCN) feature extractor (FE) to facilitate prediction with visual identification of breast thermography images to differentiate the normal from abnormal thermograms.

For the purposes of reading the present specification, some of the terms used herein are defined as follows:

- DarkCovidNet (DCN) model is an automatic detection model of COVID-19 using X-ray images, without requiring any handcrafted feature extraction techniques. DarkNet19is the modified pre-trained variant of DCN network and is employed as a feature extractor to derive relevant features from input image.

- Feature Map (FM) in a Convolutional Neural Network (CNN) is the output of a convolutional layer representing specific features in the input image or feature map. The input image is convolved with one or more filters to produce multiple feature maps. After passing through a convolutional layer, the image becomes abstracted to a feature map (FM), also called an activation map (AM). The final output from the series of dot products from the input and the filter is known as a feature map, activation map or a convolved feature.

- Feature Extractor (FE) refers to a means for extraction of relevant features such as data that is lower-dimensional than the raw data itself for example the color or the edges of an object in an image.

- Breast Thermogram (BT) is a record of thermal breast images using Thermography, also called thermal imaging, using a Digital infrared thermal imaging camera to measure the temperature of the skin on the breast's surface. In the present invention the inventors have developed a breast thermogram, also referred to as the Amrita Breast Thermogram (ABT).

- Convolutional Neural Network (CNN) is a network architecture for deep learning that learns directly from data. CNNs are particularly useful for finding patterns in images to recognize objects, classes, and categories. Convolutional Neural Network (CNN) is the extended version of artificial neural networks (ANN) which is predominantly used to extract the feature from the grid-like matrix dataset. For example, visual datasets like images or videos where data patterns play an extensive role.

The system (S) of the present invention is captured in Fig. 1, which is a flow diagram of the proposed ABN-DCN model. ABN refers to Attention Branch Network which is employed to extend a response-based visual explanation model by integrating a branch structure with an attention mechanism. ABN produces an attention map (AM) for visual explanation in feed forward propagation. ABN consists of three modules namely:
- Feature Extractor (FE)- extracts the relevant feature maps (FM) from input images.
- Attention Branch (AB)- outputs the attention location based on Class Activation Mapping (CAM) to an attention map (AM) by using an attention mechanism. CAM is a weighted activation map generated for each image. It helps to identify the region CNN is looking at while classifying an image.
- Perception Branch (PB)- receives feature maps (FM) and attention maps (AM) and outputs the probability (P) of each class.

ABN is an approach designed to improve the performance of CNN by extending an attention map (AM) for visual explanation to attention mechanism. DCN model is selected as a feature extractor (FE) because of the provision of lesser depth with fewer parameters without compromising performance accuracy. Transfer learning technique (TLT) is applied to the DCN model and fine-tuned it for thermography image classification. In order to boost the performance of DCN model, an Attention Branch Network (ABN) that combines the attention branch (AB) with DCN feature extractor is introduced. This unified unit, referred to as the ABN-DCN model, upholds an unambiguous classification of thermography image. Besides, the concurrent visualization of abnormal regions promotes diagnostic interpretability. Further, the perception branch (PB) classifies the thermographic images into normal and abnormal classes.

The ABN-DCN model is evaluated with the enhanced breast thermogram (BT) such as Amrita Breast Thermogram (ABT) dataset compiled in Indian scenario. The dataset comprises of subjects in India collaborating with the super-specialty hospital, such as Amrita Institute of Medical Science (AIMS) to enable dataset compilation for system development. The image acquisition procedure is reviewed by the ethical review committee of the said super-specialty hospital.
The procedure of capturing breast thermogram (BT) commences with the subject’s (SJ) consent and then history, if any, is accessed prior to the collection of breast images. The subject (SJ) is made to wait in an air-conditioned room for 15 min to adjust the body temperature with ambient temperature. A thermal camera (TC) is used to take breast thermography images of the subject (SJ). The distance between the subject (SJ) and the thermal camera (TC) is set as 1 m. An off the shelf thermal camera CG 640, configured with a sensitivity of less than 60mK is used to capture the breast thermograms (BT).

Fig. 2 depicts the thermal camera-based system set up for image acquisition. The images from the thermal camera (TC) are acquired with the assistance of a clinical thermographer and the validation of the collected samples is done by a clinical thermologist. The thermal camera (TC) captures and creates an image of an object by using infrared radiation emitted from the object, to produce a heat map by the process called thermal imaging. The created image represents the temperature of the object.

The interpretation of the image captured by the thermal camera (TC) is again verified using the clinical reports available in the Hospital Information System (HIS). Based on the verification, the dataset is formulated into two classes: thermograms with abnormalities and thermograms without any abnormalities (healthy subjects). Fig. 3 provides image samples in the dataset. Data augmentation techniques like rotation of 25o and horizontal flip are applied to increase the number of images in the training set for all classes to eliminate overfitting.

The procedure below discloses the CNN model for feature extraction. The DarkNet19 based DCN model used for COVID-19 detection system from chest X-ray images, is adopted as baseline model for the interpretable expert computer assisted system of the present invention. The baseline DCN model employs a smaller number of convolutional filters (CF) than DarkNet19 without compromising the performance of the model. This markedly simplifies the complexity and training time of the model. The model is retrained and fine-tuned for breast thermography ABT dataset of the invention.

Fig. 4 discloses the architecture of the proposed interpretable computer assisted expert system. The DCN feature extractor (FE) block comprises of 17 convolution layers (CL), where the Batch Normalization layer (BNL) and Leaky ReLU operations follow each layer. Each set of 3 consecutive convolution layers (CL) has identical configurations 3 times successively. The batch normalization operation has been embraced for the reduction of training time and enhancement of the model’s stability. The DCN model adopts Max Pooling (MP) to downsize an input by considering the maximum of a region selected by its filter. The image resolution is modified to 256×256 pixels to suit the requirement of input image sizes, conforming to the mentioned DCN model. The feature maps (FM) from the final convolution block of DCN model are fed to the attention branch (AB) to generate attention maps (AM) to provide diagnostic interpretability.

The ABN-DCN model is employed to provide an interpretable decision, based on the breast thermogram (BT). The Attention Branch Network (ABN) is employed to extend a response-based visual explanation model by integrating a branch structure with an attention mechanism. ABN is an approach designed to improve the performance of CNN by extending an attention map (AM) for visual explanation to attention mechanism. ABN produces an attention map (AM) for visual explanation in feed forward propagation. As mentioned in the definition clause above, the ABN comprises of three modules namely: Feature Extractor (FE), Attention Branch (AB) and Perception Branch (PB).

As depicted in Fig. 4, the CAM based attention branch (AB) is integrated after the feature extraction. In order to generate an attention map (AM), ABN uses the attention branch (AB), a top layer based on CAM that comprises of a convolution layer (CL) and Global Average Pooling (GAP). Attention map (AM) is generated using the feature map (FM) and weight at a fully connected layer after training. Hence CAM is unable to generate attention map (AM) while it is being trained. To solve this issue, ABN adopts a K × 1 × 1 convolution layer (CL) in favour of the fully connected layer, as with CAM. K is the number of categories. In feed forward processing, this K × 1 × 1 convolution layer (CL) serves as the last fully connected layer of CAM. After the K × 1 × 1 convolution layer (CL), the attention branch (AB) uses response of GAP and softmax function to output the class probability (CP). Finally, K × h × w feature map (FM) from the feature extractor (FE) is then used by the attention branch (AB) to generate an attention map (AM). A 1 × 1 × 1 convolution layer (CL) is then applied to generate 1 × h × w feature map (FM). Sigmoid function is then used to normalize the 1 × h × w feature map (FM) so that it can be utilized as the attention map (AM) for the attention mechanism.

This attention map (AM) is multiplied with feature map (FM) for highlighting the feature map (FM) at the peak of the attention map (AM) as per Eq.1. The generated feature map ?? ' ?? (???? ) with softmax function in the perception branch (PB) predicts the probability of each class (CP) for the binary classification of thermography images. After obtaining the attention maps (AM) and feature maps (FM), perception branch (PB) outputs the final probability (P) for each class (CP). First, Attention mechanism applies the attention map (AM) to the feature map (FM). In this case, ???? (????) is the feature map (FM) at the feature extractor (FE); ??(???? ) is an attention map; and ?? ' ?? (???? ) is the output of the attention mechanism, where c|1, . . . , C is the index of the channel.

A simple dot-product of the attention maps (AM) and feature maps (FM) at a particular channel c constitutes Eq 1. ABN can be trained in an end-to-end manner using losses at both the attention branch (AB) and the perception branch (PB). According to Eq. 2, the training loss function ??(???? ) is a simple sum of losses at both branches, as expressed by Eq. 2.

Here, ????????(???? ) denotes training loss of input image ???? at the attention branch (AB), and ????????(???? ) denotes training loss at the perception branch (PB). For the training of BreaKHis dataset, selected training loss function is selected for attention branch (AB) as weighted cross entropy. Sample weighting (SW) is applied to address the model bias while using the highly imbalanced ABT dataset. In order to address the probability score, cross entropy loss and softmax function are used in the perception branch.

CLASSIFICATION OF THERMOGRAPHY IMAGES BASED ON ABN-DCN MODEL:
The ABN-DCN model is now applied for classification of breast thermography images. The Enhanced ABT dataset used in this case consists of 68 Abnormal images and 263 normal images.

For the experimental verification of the ABN-DCN model, an instance of the Tesla K-80 GPU was availed of, provided by Google Collaboratory. The images in the dataset are resized to 256 x 256 which is suitable for the proposed DCN architecture. For the training, 500 epochs are set with batch size of 32. A learning rate of 0.1 is adopted for controlling the speed of training and adopted Stochastic Gradient Descent (SGD) optimizer to converge the training.

The model is trained and validated with a new dataset i.e., Amrita Breast Thermogram (ABT) dataset. The classification performance of this model was evaluated using popular benchmark metrics-accuracy, precision, recall, and F1 score. Transfer learning technique (TLT) was applied on these models and fine-tuned for thermography images in the ABT dataset.

Comparative appraisals of ABN-DCN model with other existing low-footprint CNN models such as baseline DCN model, VGG16, EfficientNet, MobileNet are tabulated in Table 1. It is noticed that integration of the attention map (AM) with the baseline DCN model forming the ABN-DCN model greatly increases the classification performance compared to other CNN models.

As depicted in Table 1, the model exhibited 96.7 % accuracy using ABT dataset, demonstrating comparable performance in binary classification (BC) of thermography images. Moreover, introducing Attention Branch Network (ABN) boosted the performance of the baseline DarkNet19 CNN (DCN) model by 8% on ABT dataset. Therefore, the baseline DCN model convincingly outperforms conventional low-footprint CNN models.

Notably, the ABN-DCN model outperforms the baseline DCN model, showcasing a significant enhancement of 5.5% in terms of accuracy. The incorporation of sample weighting (SW) through the application of a weighted cross-entropy loss within the attention branch (AB) plays a pivotal role in mitigating model bias when dealing with the imbalanced ABT dataset. This strategic enhancement contributes to a remarkable accuracy improvement of 2.39%, culminating in an impressive accuracy level of 96.7%. To verify the generalizability of the ABN-DCN model, the performance of the model is evaluated using publicly available DMR-IR dataset which shows a remarkable performance of 95% with this dataset.

Table 1: Performance comparison of ABN-DCN model with other low foot
print CNN models.

Model Accuracy Precision Recall F1 Score
VGG16 92.85 81.2 95.7 87.8
MobileNet 86.33 79.84 77.3 78.5
EfficientNet 85.3 70.9 78.7 74.5
DCN Model 89.5 90.3 93.3 91.8
ABN DCN model 94.44 94.8 97.2 95.9
ABN DCN Model with sample weighting 96.7 95.8 98.3 97.1
The decision support system which combines ABN with the modified DarkNet19 (DCN) model, not only provides cutting-edge performance but also diagnostic interpretability, allowing doctors to identify the region of focus from which the model makes its decision making. This can potentially lead to improved clinical translation of deep learning models for clinical decision support.

Attention map visualization with ABN-DCN model
Fig. 5 displays the attention map (AM) generated by ABN-DCN model using images drawn from the Amrita Breast Thermogram (ABT) dataset. The attention map (AM) clearly demonstrates that the ABN-DCN model precisely focuses on the breast area in the thermograms for decision-making. It is observed that features from the whole breast area are considered, analyzing the thermal symmetry in the breast for the classification of thermal images. The results presented here are encouraging and confirm the potential of thermography to assist in detecting breast cancer in its early stages.

The captured breast thermograms (BT), the patient’s details (Basic information about the patient, consent form with patient’s signature, and patient history form), and the breast screening results provided by the system are updated to the doctor through various means including manually or automatically, for example through REDCap software for further diagnosis. The created REDCap account can be accessed in the tertiary hospitals/ doctor’s smartphone or PC with the login credentials, allowing doctors to see the entered patient data that helps the doctor conduct further analysis and diagnosis.
, Claims:1. A system (S) for obtaining clinical inference from thermograms, said system comprising:
ABN-DCN integration model, wherein:
- the DarkCovidNet (DCN) is modified as feature extractor (FE) to extract discriminative features from the Region of Interest (ROI) of thermography images to generate feature maps (FM),
- the Attention Branch Network (ABN) is integrated with DarkCovidNet (DCN) feature extractor (FE) to form the ABN-DCN model for binary classification of thermography images, said ABN comprises:
o Attention Branch (AB) which utilizes attention mechanism for outputting attention location to generate an attention map (AM), based on Class Activation Mapping (CAM),
o Perception Branch (PB) module to classify said thermography images,
characterized in that:
- said feature maps (FM) from the final convolution block of DCN model are fed to the said attention branch (AB) to generate attention maps (AM) enabling diagnostic interpretability,
- intrinsic features of thermography images are combined with prediction for binary classification of said thermography images, to recognize abnormality from thermograms.

2. The system (S) as claimed in claim 1, wherein Breast Thermogram (BT) comprises of breast thermography images based on image acquisition using thermal camera (TC) and associated software, and clinical information of the subject (SJ) is validated by the Hospital Information System (HIS),

3. The system (S) as claimed in claim 1, wherein the DarkCovidNet (DCN) is used as a modified version of DarkNet19 to capture and learn the relevant features from input histopathology images by applying transfer learning technique (TLT).

4. The system (S) as claimed in claim 1, wherein the DCN feature extractor (FE) comprises convolutional layers (CL) where the Batch Normalization Layer (BNL) and LeakyReLU operations follow each layer.

5. The system (S) as claimed in claim 1, wherein the Attention Branch Network (ABN) introduces an attention branch (AB) using response based visual explanation to emphasize the region of focus of the CNN model.

6. The system (S) as claimed in claim 1, wherein the said attention branch (AB) uses a convolutional layer of DarkCovidNet (DCN) and Global Average Pooling (GAP) to generate a heatmap to identify the region (ROI) of interest in thermography images.

7. The system (S) as claimed in claim 6, wherein, the convolutional layers are 17 in number whereas the Max Pooling layers are 5 in number.

8. The system (S) as claimed in claim 1, wherein the perception branch (PB) receives feature maps (FM) and attention maps (AM) and output the probability (P) of each class of thermography images.

9. The system as claimed in claim 8, wherein the probability (P) of classes is used for binary classification of thermography images into normal or abnormal.

10. The system (S) as claimed in claim 1, wherein, sample weighting technique is implemented to mitigate dataset imbalance in thermography images, thereby alleviating model bias favoring the dominant class.
11. A method of clinical inference of thermography images of the subject (SJ) to identify abnormality from said thermography images by the system (S) as claimed in claim 1, comprising the steps of:
- capturing breast thermography images by thermal camera and validating the same from clinical data obtained from the Hospital Information System (HIS),
- extracting features from thermography images by DCN feature extractor to generate feature map (FM),
- inputting the said feature maps (FM) from the final convolutional block of DCN into the attention branch (AB), to generate attention map (AM),
- multiplying attention map (AM) with feature map (FM) for highlighting feature map (FM) at the peak of attention map (AM),
- feeding of feature map (FM) and attention map (AM) to the perception branch (PB),
- computing binary classification of thermography images by outputting probability (P) to each class of the said images,
- updating the captured breast thermograms (BT), the patient’s details (Basic information about the patient, consent form with patient’s signature, and patient history form), and the breast screening results provided by the said system (S) to the doctor through various means including manually or through automatic software such as REDCap software.