Breast cancer detection system using a-mode ultrasound sensor

Abstract

The present invention relates to a medical diagnostic system configured to automatically detect breast cancer using an A-mode ultrasound sensor, on the basis of a reflection signal from inside a human body. More specifically, the present invention relates to a breast cancer detection system using an A-mode ultrasound sensor, the system combining low-cost and compact A-mode ultrasound technology with an artificial intelligence-based analysis algorithm to enable real-time breast cancer self-diagnosis and convenient diagnosis of breast cancer at a primary healthcare institution.

Description

Breast cancer detection system utilizing A-MODE ultrasound sensors

[0001] The present invention relates to a medical diagnostic system that automatically detects breast cancer based on reflected signals inside the human body using an A-mode ultrasound sensor. More specifically, the invention relates to a breast cancer detection system using an A-mode ultrasound sensor that enables real-time breast cancer self-diagnosis and simple breast cancer diagnosis at primary medical institutions by combining low-cost and miniaturized A-mode ultrasound technology with an artificial intelligence-based analysis algorithm.

[0002] Recently, the incidence of breast cancer has been continuously increasing both domestically and internationally, and early detection, particularly among middle-aged women aged 40 and older, is emerging as a critical public health issue. Due to various factors such as the Westernization of dietary habits, late marriages and declining birth rates, genetic factors, and increased life expectancy, breast cancer is no longer limited to specific high-risk groups but is becoming a more widespread concern across a broader age range. These changes imply that breast cancer prevention and early diagnosis must target all women, not just specific demographics, and consequently, the need for routinely accessible regular screening systems is growing.

[0003] However, current early breast cancer detection methods primarily consist of mammography and high-resolution B-mode ultrasound, although some attempts are made to identify breast cancer through self-examination. Nevertheless, self-examination has significantly low accuracy, and B-mode ultrasound equipment is expensive medical device that generates images using thousands of sensors; it faces limitations such as high costs amounting to hundreds of millions of won and the mandatory interpretation by a specialist radiologist. Consequently, since it is practically impossible for the general public to purchase such expensive equipment or interpret the images without medical expertise, accessibility to early diagnosis is significantly reduced.

[0004] To overcome these structural and economic limitations, the development of a breast cancer detection system that is more cost-effective and easy for non-experts to use is essential. Accordingly, there is an urgent need to develop a new technology-based breast cancer diagnostic system that is easy for anyone to use while reliably determining the presence or absence of breast cancer.

[0005] Accordingly, the technical problem of the present invention is conceived from this point, and the objective of the present invention is to provide a breast cancer detection system utilizing an A-mode ultrasound sensor that utilizes an A-mode ultrasound method without requiring a complex B-mode image generation device.

[0006] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasonic sensor in which the transducer and signal processing circuit are simply configured based on a single-channel transmission and reception structure.

[0007] In addition, it provides a breast cancer detection system utilizing an A-mode ultrasound sensor that can determine the presence of breast cancer solely through amplitude-based signal analysis, without undergoing a complex image interpretation process.

[0008] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasound sensor that simultaneously learns local abnormal signals and global structural distortions by applying CNN and ViT-based artificial intelligence models in parallel.

[0009] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasound sensor in which ultrasound signals collected through an A-mode probe are digitized and transmitted to a user terminal in real time, thereby providing the analysis results to the user immediately.

[0010] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasound sensor in which data augmentation based on user operation variability (e.g., scan speed and pressure) is performed through the VSPA algorithm.

[0011] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasound sensor that performs region-level prediction by integrating not only a single image but also multiple A-mode data collected from the same patient.

[0012] In addition, it provides a breast cancer detection system utilizing an A-mode ultrasound sensor that visually highlights suspicious sections within the waveform via a user terminal and explains the basis for the artificial intelligence's judgment in natural language.

[0013] In addition, the invention provides a breast cancer detection system utilizing an A-mode ultrasound sensor designed for flexible use in various user terminal environments, such as mobile apps, hospital web clients, and medical tablets.

[0014] To realize the objective of the present invention, the present invention provides a human breast cancer detection system utilizing an A-mode ultrasonic sensor, comprising: an input unit that receives data from the outside into the system; an output unit that outputs result data to the outside; a communication unit that transmits and receives data between the inside and outside of the system; a storage unit that stores data generated by the system; a control unit that controls the operation of the system; and a memory unit that executes various programs and stores data accordingly. The memory unit comprises: a data collection unit that transmits an A-mode ultrasonic signal to the human body and receives an ultrasonic signal reflected from within the human body; a data preprocessing unit that preprocesses data collected from the data collection unit; an artificial intelligence analysis unit that determines the presence of breast cancer based on the preprocessed data; and a user interface unit that provides the analysis results of the artificial intelligence analysis unit to a user.

[0015] According to the breast cancer detection system utilizing a real-time A-mode ultrasound sensor according to the embodiments of the present invention, the A-mode ultrasound method does not require a complex B-mode image generating device, so the manufacturing cost of the entire system can be drastically reduced.

[0016] In addition, since the transducer and signal processing circuits are simply configured based on a single-channel transmission and reception structure, the system can be made lightweight and easy to implement as a portable diagnostic device.

[0017] In addition, since the presence of breast cancer can be determined solely through amplitude-based signal analysis without undergoing a complex image interpretation process, general users can perform self-diagnosis without any separate medical expertise.

[0018] In addition, by applying CNN and ViT-based artificial intelligence models in parallel, local abnormal signals and global structural distortions can be learned simultaneously, which can significantly improve the accuracy and sensitivity of breast cancer detection.

[0019] In addition, ultrasound signals collected through the A-mode probe are digitized and transmitted to the user terminal in real time, and analysis results are provided immediately, enabling rapid clinical response.

[0020] In addition, data augmentation based on user operation variability (e.g., scan speed and pressure) is performed through the VSPA algorithm, improving the generalization performance and robustness of the AI ​​model.

[0021] In addition, by integrating multiple A-mode data collected from the same patient as well as a single image to perform region-level prediction, the misdiagnosis rate can be reduced and diagnostic reliability increased.

[0022] In addition, user trust and understanding can be improved by visually highlighting suspicious sections within the waveform through the user terminal and explaining the basis of the artificial intelligence's judgment in natural language.

[0023] In addition, it is designed to be used flexibly in various user terminal environments, such as mobile apps, hospital web clients, and medical tablets, providing high practicality.

[0024] However, the effects of the present invention are not limited to the above effects and may be extended in various ways without departing from the spirit and scope of the present invention.

[0025] FIG. 1 is a block diagram of a breast cancer detection system utilizing an A-mode ultrasonic sensor according to an embodiment of the present invention.

[0026] FIG. 2 is a diagram showing the distinction between A-mode ultrasound and B-mode ultrasound according to an embodiment of the present invention.

[0027] FIG. 3 is a diagram showing the process of transmitting and receiving ultrasonic data in a data collection unit according to an embodiment of the present invention.

[0028] FIG. 4 is a diagram showing the process of preprocessing ultrasonic data received in a data preprocessing unit according to an embodiment of the present invention.

[0029] FIG. 5 is a diagram showing the process of detecting breast cancer in the human body using an artificial intelligence model in an artificial intelligence analysis unit according to an embodiment of the present invention.

[0030] FIG. 6 is a diagram showing a method for evaluating an artificial intelligence model used in an artificial intelligence analysis unit according to an embodiment of the present invention.

[0031] FIG. 7 is a diagram showing that a breast cancer detection result according to an embodiment of the present invention is provided through a user's terminal.

[0032] FIG. 8 is a diagram showing the results of evaluating the performance of an artificial intelligence model according to an embodiment of the present invention.

[0033] Hereinafter, preferred embodiments of the present invention will be described in more detail with reference to the drawings.

[0034] The present invention is capable of various modifications and may take various forms, and specific embodiments are illustrated in the drawings and described in detail in the text. However, this is not intended to limit the invention to the specific disclosed forms, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention.

[0035] FIG. 1 is a block diagram of a breast cancer detection system utilizing an A-mode ultrasonic sensor according to an embodiment of the present invention.

[0036] Referring to FIG. 1, external input is collected through an input unit (110), which converts a biosignal (reflected ultrasound) obtained through a probe into an electrical signal and transmits it to an internal module. The collected signal is stored in a storage unit (140), and at the same time, a bidirectional connection is maintained wirelessly with a user terminal (20) through a communication unit (130). The entire system operates under the control of a control unit (150) and coordinates the signal flow and operation sequence between each component.

[0037] The core analysis function of the breast cancer detection system (100) utilizing an A-mode ultrasound sensor is implemented through four main modules included in the memory unit (160). The data collection unit (170) receives raw ultrasound data input from the probe and is responsible for the initial stage of preprocessing, which converts the data into a digital signal. The subsequent data preprocessing unit (180) processes the data into a data optimized for model learning and inference by applying the VSPA (Variable Speed ​​and Pressure Augmentation) algorithm to correct deviations such as scan speed and pressure caused by user operation. Subsequently, the artificial intelligence analysis unit (190) learns the structural abnormalities of the A-mode ultrasound waveform by utilizing CNN-based ResNet50 and ViT (Vision Transformer)-based ViT-B / 16 models in parallel, and automatically classifies whether or not there is breast cancer.

[0038] The analyzed results are visually provided to the user through the user interface unit (200). The user interface unit (200) does not merely display the diagnosis results of the artificial intelligence model in the form of simple numbers or sentences, but enables intuitive recognition by visually highlighting the areas suspected of breast cancer within the image. In particular, to enable the interpretation of the decision-making process of the artificial intelligence model, interpretability information in the form of a heatmap applying the Grad-CAM (Gradient-weighted Class Activation Mapping) algorithm is provided, thereby supporting the user in understanding the basis from which the result was derived, which contributes to improving user trust. The interpretability in the form of a heatmap visually indicates which image areas the model focused on when determining the presence of breast cancer, and presents the basis for the intuitive judgment by displaying areas in red as their importance increases. In addition, the results are designed to be flexibly optimized and output for various user environments, such as mobile applications, hospital web clients, and medical tablets, and the diagnostic history is stored and configured with a structure that has scalability to be linked with the electronic medical record (EMR) system of a medical institution in the future.

[0039] FIG. 2 is a diagram showing the distinction between A-mode ultrasound and B-mode ultrasound according to an embodiment of the present invention.

[0040] Referring to Fig. 2, A-mode ultrasound is the most basic form of ultrasound diagnostic technology, which represents the amplitude of acoustic signals reflected back from tissue boundaries within the human body as a one-dimensional waveform with respect to the time or distance axis. This method utilizes the characteristic that sound waves radiated in a single direction by an ultrasound transducer are reflected when they reach a boundary surface within human tissue that has a difference in density or acoustic impedance. The arrival time of each reflected wave represents the reflection location (depth), and the amplitude represents the reflection intensity. The output result is presented in the form of an amplitude profile rather than an image, which allows for the quantitative determination of the presence or absence of boundaries between tissues.

[0041] A-mode ultrasound offers distinct advantages over B-mode ultrasound in terms of structure, cost-effectiveness, and applicability. The most prominent features are the structural simplicity and miniaturization potential of the equipment. Based on a single-channel ultrasound transmission and reception structure, A-mode can be implemented without complex array transducers or image generation hardware, which is advantageous for miniaturizing and lightweighting the entire system. These characteristics significantly expand the applicability of A-mode ultrasound technology, particularly in the field of personalized medical devices such as wearable devices and portable self-diagnostic equipment.

[0042] In addition, A-mode ultrasound is highly efficient in terms of manufacturing and operating costs. While B-mode ultrasound, which generates high-resolution images, requires expensive imaging engines, high-speed digital processing units, and complex interfaces and is limited to stationary equipment centered on medical institutions, A-mode can operate with only simple transmit / receive circuits and basic signal processing processors. Due to these characteristics, A-mode-based devices can be implemented at a cost level that is tens of times lower than that of B-mode, and can contribute significantly to expanding access to public healthcare and medical services in developing countries by enabling widespread distribution at low cost.

[0043] In terms of technical functionality, A-mode is differentiated by its ability to perform signal-based quantitative analysis rather than image-based interpretation. By directly quantifying the amplitude and time delay of ultrasound reflection signals to form an A-mode waveform, and combining this with AI-based pattern recognition and classification algorithms, pathological characteristics such as the presence or absence of breast cancer can be automatically determined. This approach enables an autonomous diagnostic environment that eliminates the need for specialized radiology interpretation and can be expanded into a system capable of providing meaningful diagnostic information in real time to patients or primary care institutions.

[0044] FIG. 3 is a diagram showing the process of transmitting and receiving ultrasonic data in a data collection unit according to an embodiment of the present invention.

[0045] Referring to FIG. 3, the data acquisition unit (170) acquires a bio-signal inside the human body through an A-mode probe (a sensor device that transmits an ultrasonic signal to the human body and receives an ultrasonic signal reflected from inside the human body) and performs a series of processes to digitize and process it so that artificial intelligence analysis is possible. The data acquisition unit (170) performs operations including the entire signal flow from the generation of an ultrasonic transmission signal, reception of a reflected wave, amplification of an electrical signal, analog-to-digital conversion, signal processing, wireless transmission, and reception to a user terminal, and forms the basis for the accuracy and real-time diagnostic function of the entire system.

[0046] First, the transmitting processor generates low-voltage and low-current pulses at a constant period, and these pulses are transmitted to the ultrasound pulse driver. The ultrasound pulse driver amplifies the input signal into high voltage and high current to convert it into a signal capable of driving the ultrasound transducer, and the high-voltage pulse is subsequently applied to the ultrasound transducer through a transmit / receive switch (T / R switch). In transmit mode, the T / R switch supplies the high-voltage driving signal to the transducer, and in receive mode, it converts the ultrasound signal reflected from the transducer into an electrical signal and transmits it to the lower circuit.

[0047] An ultrasonic transducer receives high-voltage electrical pulses and converts them into mechanical ultrasound, which propagates through the human skin into internal tissues. The ultrasound is reflected at the boundaries where the acoustic impedance within the tissue changes, and the reflected waves return to the transducer to be converted into electrical signals. Since the acquired electrical signal is a weak analog signal with very low amplitude, it is sufficiently amplified through a low-noise high-gain amplifier. The amplified analog signal is then converted into a digital signal via an analog-to-digital converter (ADC).

[0048] The digitized signal is subsequently post-processed by a receiving signal processing processor. In the signal processing stage, (1) digital filtering to remove high-frequency or low-frequency noise, (2) envelope detection to extract the contour of the reflected signal to secure a clear signal shape, and (3) logarithmic compression to compress the amplitude range of the signal to improve analysis and visualization efficiency are performed sequentially. This completes one-dimensional A-mode waveform data (Amplitude-mode Signal Profile) with an amplitude distribution over time.

[0049] The completed A-mode waveform is stored in the system's built-in memory or transmitted in real-time to an external user terminal (mobile device) via a Bluetooth communication module. Based on the received data, the mobile device visualizes the ultrasound waveform and automatically determines the presence of breast cancer within the body using an integrated artificial intelligence breast cancer detection model. As a result, users can check diagnostic results in real-time without specialized knowledge, and the device offers high portability and practicality based on a low-power, miniaturized design.

[0050] FIG. 4 is a diagram showing the process of preprocessing ultrasonic data received in a data preprocessing unit according to an embodiment of the present invention.

[0051] Referring to Fig. 4, unlike B-mode ultrasound, A-mode ultrasound does not provide a two-dimensional image and generates only one-dimensional waveform data that represents the amplitude of the ultrasound reflection signal with respect to the time or distance axis. Due to this characteristic, the waveform shape and data quality of A-mode data fluctuate sensitively depending on the user's scan speed and probe pressure. In other words, even when scanning the same area, the shape of the acquired signal may vary depending on how quickly the user moves the ultrasound sensor or how hard they press it. Since inconsistency caused by such manipulation deviations can impede the generalization performance of the trained artificial intelligence model, a systematic correction method is absolutely necessary.

[0052] To address this, the data preprocessing unit (180) preprocesses the acquired data using the VSPA (Variable Speed ​​and Pressure Augmentation) algorithm. VSPA is a data augmentation technique that simultaneously improves the sensitivity and robustness of an artificial intelligence model by augmenting the dataset through artificial simulation of variations in scan speed and pressure that may actually occur in a clinical environment. The algorithm is composed of three main types of augmentation methods, which are horizontal augmentation (scan speed variation), vertical augmentation (pressure intensity variation), and local range compression.

[0053] First, Horizontal Augmentation is a method that simulates changes in user scanning speed. The input image has a height H and a width W, and mathematically I R H*W It is defined as follows. For this image, a resizing factor α (where α is experimentally selected) is randomly sampled from a range between 0.7 and 1.5 and applied as follows.

[0054] When -α < 1: Reflects the situation where the user scanned quickly, the image is compressed by a factor of α along the horizontal axis, and the reduced width W'=α*W(I resized R H*W' The insufficient area (empty space) is compensated for by zero padding. For example, when α is 0.8, the original width W is reduced to 80% of its size, W', and the remaining area (W- W') that is insufficient due to compression is filled with zeros through zero padding.

[0055] - When α > 1: This reflects the situation where the user scans slowly. The image is expanded by a factor of α along the horizontal axis, and then center cropping is performed to fit the model input size. For example, if α is 1.3, an image is generated that is 30% wider than the original width. Subsequently, the expanded image undergoes center cropping to cut off both edges, leaving only the central area, in order to fit the fixed input size of the artificial intelligence model.

[0056] Second, vertical augmentation is a method that reflects signal deformation caused by changes in probe pressure. The application method is similar to horizontal augmentation, but the reference axis is the vertical axis (H), and the resizing factor α is applied only within the range where α < 1, assuming only situations where pressure increases. In actual clinical situations, pressure increases more frequently than decreases, so only cases corresponding to compression are reflected in augmentation. Even in this case, the empty area (empty space) corresponding to the reduced height is compensated for through zero padding. For example, if the vertical length of the original image is 100 pixels and the resizing factor α is set to 0.8, the image is compressed by 0.8 times in the vertical direction and reduced to 80 pixels. Afterwards, the reduced 20 pixels of empty space are filled with zero values ​​evenly at the top and bottom through zero padding, thereby maintaining the original input size of 100 pixels.

[0057] Third, Local Region Compression is an enhancement method designed to reflect local variations (e.g., variations caused by slippage due to changes in skin surface texture in elderly patients, variations caused by uneven surfaces in cases of large tumors), such as pressure changes or hand tremors applied only to specific sections during user operation. By selecting an arbitrary section in the horizontal direction within the image, the length W of this section S = x e - x s A resizing factor α in the range of 0.1 to 0.5 is applied to . The corresponding region is W ' S = α * W S It decreases to, and the remaining reduced part W S - W ' S The entire image shape is maintained by zero-padding along the horizontal axis. This method can effectively simulate real-world diagnostic environments, particularly where minute transducer movements affect only specific signal segments. For example, assuming that a segment from the 30th pixel to the 50th pixel is selected from an image of 100 pixels along the horizontal direction (time axis) of an A-mode ultrasound waveform, the length of that segment is W S = 50 - 30 = 20 pixels, and if the resizing factor α is set to 0.5, this section is compressed by half in the horizontal direction, W ' S = 0.5 * 20 = 10 pixels. The remaining 10 pixels are filled with zero padding to maintain the continuity of the waveform and the image structure.

[0058] As described above, the VSPA algorithm precisely reflects the one-dimensional characteristics of A-mode ultrasound and user operation variability in a clinical environment, and augments training data so that the artificial intelligence model can maintain high sensitivity and robustness even under various input conditions.

[0059] FIG. 5 is a diagram showing the process of detecting breast cancer using an artificial intelligence model in an artificial intelligence analysis unit according to an embodiment of the present invention, and FIG. 6 is a diagram showing the method of evaluating an artificial intelligence model used in an artificial intelligence analysis unit according to an embodiment of the present invention.

[0060] Referring to FIGS. 5 and 6, the artificial intelligence analysis unit (190) automatically determines the presence of breast cancer based on a one-dimensional A-mode waveform (Amplitude-mode Signal Profile) collected along a time or distance axis, which contains amplitude information of ultrasound reflected back from the internal tissue boundary of the human body. The artificial intelligence analysis unit (190) utilizes the structural characteristics of A-mode ultrasound data and abnormalities of reflection patterns as learning data, and in particular, quantifies the difference in reflection signal characteristics between normal tissue and breast cancer tissue and uses this as a classification criterion.

[0061] The artificial intelligence analysis unit (190) utilizes a CNN-based ResNet50 and a Vision Transformer (ViT)-based ViT-B / 16 model in parallel, and each model reflects a different analysis perspective of the A-mode waveform. First, ResNet50 analyzes local amplitude deviations within the waveform, that is, abnormal echo magnitude shifts occurring at tissue boundaries. For example, in normal tissue, the reflection intensity changes gradually, but if breast cancer is present, there is a tendency for a sudden amplitude spike or drop to occur in that area. In addition, heterogeneous boundary reflections at the breast cancer boundary are also detected, which are related to the heterogeneity of the tissue.

[0062] On the other hand, the ViT-B / 16 model converts A-mode ultrasound waveforms into a two-dimensional image form and analyzes them by dividing them into small regions (patches). Each patch represents a part of the entire waveform, and the model processes these patches individually and identifies the structure of the entire waveform by comparing their relationships. The technique used in this process is called Multi-head Self-Attention, which is a method that analyzes how each patch relates to other patches from multiple perspectives simultaneously.

[0063] Through this, the ViT-B / 16 model detects abnormal structures (structural distortion, Global Pattern Irregularity) that differ from the typical reflection signal pattern throughout the waveform, or parts where the flow between signals that should be connected is broken or awkwardly connected (Contextual Discontinuity). For example, while reflection signals in normal tissue show a consistent pattern, in the case of breast cancer, such patterns appear to be misaligned or broken, and ViT-B / 16 detects these signs of anomalies in the overall flow.

[0064] Breast cancer tissue generally exhibits weak pattern connectivity with adjacent reflection locations and positionally incoherent structures; ViT identifies breast cancer regions by detecting these global pattern distortions.

[0065] The two models learn the same A-mode data from different perspectives and ultimately determine the presence of breast cancer based on the following consistent breast cancer diagnosis criteria.

[0066] First, the first criterion is the presence of Abnormal Amplitude Zones, which refer to zones where excessive or rapid amplitude changes occur compared to normal tissue at specific times or distances. In typical normal tissue, ultrasound reflection signals show a constant amplitude distribution, but in the presence of breast cancer, there is a tendency for abnormally high echo intensity or rapid signal degradation to appear in the affected area.

[0067] The second criterion is irregular echo spacing. This refers to a phenomenon where the spacing between adjacent reflected signals within a waveform is inconsistent, with signals appearing either densely packed or excessively spread out in certain sections. While normal tissues exhibit a structure where reflected signals repeat at regular intervals, breast cancer tissue shows atypical variations in these spacing due to the heterogeneity of its internal structure.

[0068] The third criterion is sharpness of the boundary and sharp edge reflections. This refers to a phenomenon where reflection boundaries are sharply cut at specific points within the A-mode waveform, or where the reflection intensity changes abruptly at a single boundary. Since the boundaries of breast cancer generally exhibit rapid density changes compared to normal tissue, they are characterized by very distinct and steep reflection differences when ultrasound passes through these boundaries.

[0069] Finally, diminished structural coherence refers to the loss of contextual continuity between amplitudes and reflection locations observed across the entire A-mode waveform. In other words, in breast cancer tissue, this manifests as a disruption of the overall rhythm, pattern, and phase structure of the waveform, resulting in a breakdown of the consistent structural flow with normal tissue signals.

[0070] The model trained based on these criteria is configured not only to predict at the image-level, but also to perform a comprehensive prediction at the region-level by integrating multiple A-mode waveform data collected from the same patient. In the above process, a soft voting technique is applied, and the predicted probability values ​​(e.g., probability of breast cancer) calculated by the artificial intelligence model for each individual waveform are averaged to derive a final diagnosis for the entire patient. This method reduces the possibility of misdiagnosis that may occur in a single image and provides a more stable and reliable judgment result by comprehensively considering multiple waveform data. The final diagnosis result is output to the user in visual and numerical forms through the user interface unit (200).

[0071] The user interface section (200) includes not only simple diagnostic results but also a judgment basis explanation function (Interpretability Text Module) that visually highlights waveform sections that had a decisive influence on the judgment and explains in natural language what signal characteristics the artificial intelligence classified breast cancer based on. For example, an explanation such as "a rapid increase in amplitude was confirmed in the 15mm section, and the possibility of breast cancer was judged to be high due to abnormal reflection characteristics" is provided so that the user can trust and understand the results.

[0072] In addition, the above models do not use fixed static data, but rather introduce a data augmentation technique applied in real-time at the time of model training to automatically generate training data that simulates changes in scan conditions in an actual clinical environment. In particular, the Variable Speed ​​and Pressure Augmentation (VSPA) algorithm proposed in the present invention is a unique augmentation method that reflects data imbalances that may occur due to deviations in user scan speed and probe pressure, and reflects detailed operation scenarios such as horizontal and vertical resizing and local range compression. As a result, the model learns A-mode data under various conditions even for the same lesion, thereby significantly improving the model's sensitivity and robustness.

[0073] In addition, after training the model based on the expanded dataset, the 5-fold cross-validation method is adopted as a verification procedure for objective performance evaluation. The 5-fold cross-validation method divides the entire dataset into five parts (folds), sequentially designating one fold for validation and using the remaining four for training. Through this process, trained models (M1 to M5) are generated for each fold, and these models perform inference on an independent test set; the mean performance of the models is then calculated by averaging the inference results of all models. The above structure enables the objective evaluation of the model's overall generalization performance while minimizing bias to the training distribution.

[0074] The model's prediction results are quantified at two evaluation criteria levels. The first is image-level evaluation, where predictions are performed individually for each A-mode image, and indicators such as AUC (Area Under Curve), Accuracy, and Sensitivity are calculated based on these results. The second is region-level evaluation (grouping and evaluating data by region when there are multiple images of normal and tumor regions for a single patient). Multiple A-mode data collected for each patient—both normal and tumor regions—are grouped by region, and a comprehensive diagnostic result is derived by applying soft voting to the prediction results of each image. This evaluation method reflects a process where judgments are made based on multiple data, rather than relying on a single image as is done in actual clinical diagnosis.

[0075] FIG. 7 is a diagram showing that a breast cancer detection result according to an embodiment of the present invention is provided through a user's terminal.

[0076] Referring to FIG. 7, the analysis results of the artificial intelligence analysis unit (190) are provided to the user intuitively and reliably through the user interface unit (200). The user interface unit (200) visualizes the artificial intelligence identification results for the collected A-mode ultrasound data and provides them to the user by comprehensively organizing them at the image-level or region-level, thereby ensuring both transparency and practicality in the diagnostic process.

[0077] The user interface section (200) is designed so that not only medical professionals but also general users can understand and utilize the results, and includes an information visualization module, a result explanation function, and a result expression differentiation function for each user type.

[0078] First, the user interface unit (200) visualizes one-dimensional amplitude waveform data obtained through an A-mode ultrasound sensor as a time-axis-based graph, and specific reflection locations suspected of breast cancer are highlighted with colors or markers based on the results of artificial intelligence analysis. Through this, the user can clearly identify which section within the waveform has a pathological abnormal signal. In addition, the prediction result at the image unit is indicated as 'positive / negative', and the reliability (prediction probability score) is displayed together, allowing for quantitative verification of the strength or uncertainty of the diagnosis.

[0079] In addition, when multiple A-mode images are collected for a single subject, the user interface unit (200) provides not only the results for each image but also the final diagnosis results at the region level to which a soft voting algorithm is applied. In the above process, the prediction results are expressed in the form of a bar chart, gauge graphic, or text summary, and are provided to the user in the form of 'suspected breast cancer (92.5% confidence)' or 'normal findings'. In particular, in the medical professional mode, quantitative indicators such as AUC, sensitivity, specificity, PPV, and NPV are provided together, while in the general user mode, these are omitted and summarized mainly with interpretable sentence-type explanations and warning messages.

[0080] Additionally, the user interface section (200) includes an interpretability module that explains the basis for the model judgment regarding the result, and automatically analyzes the waveform section that served as the main criterion for the prediction to provide explanatory information such as, "High-amplitude reflection signals were observed in the 5mm to 10mm section, indicating a high probability of breast cancer." This function contributes to helping the user understand the reason for the prediction beyond the simple result value and to increase diagnostic reliability. The interface layout and the level of output information are automatically adjusted according to the user type or environment, and this can be flexibly applied to various types of terminals, such as user mobile apps, medical tablets, and hospital web clients.

[0081] Finally, the user interface section (200) also includes a result history storage and analysis function. The user can view past diagnostic results by date or visually check the trend of changes in the A-mode waveform through time series analysis, and the result report can be configured to be exported as a PDF or standard medical information format to be linked with a medical institution or linked to an EMR.

[0082] FIG. 8 is a diagram showing the results of evaluating the performance of an artificial intelligence model according to an embodiment of the present invention.

[0083] Referring to FIG. 8, the present invention applies a 5-fold cross-validation method to objectively and reliably evaluate the generalization performance of an artificial intelligence model. This method involves dividing the entire training data into five subsets (folds), sequentially designating each fold as a validation set, and using the remaining four folds as training sets to repeat training and validation a total of five times. By averaging the performance indicators calculated for each fold to arrive at the final model's performance, performance deviations due to data splitting can be minimized, and the overall stability and reliability of the model can be ensured.

[0084] The following representative classification performance metrics are used to evaluate the quantitative performance of a model. The Area Under the Receiver Operating Characteristic Curve (AUC) is an indicator representing the model's discrimination performance across classification thresholds; a value closer to 1 indicates an ideal classification model. Accuracy is calculated as the proportion of actual correct answers among all predictions. Sensitivity represents the proportion of cases correctly predicted as positive for breast cancer and serves as a key indicator for early diagnosis. Specificity is the proportion of cases correctly predicted as negative when they are actually negative (normal), and is used as a reliability metric to prevent overdiagnosis. Positive Predictive Value (PPV) and Negative Predictive Value (NPV) quantify the proportion of actual results among those predicted as positive or negative by the model, respectively.

[0085] To quantitatively evaluate the generalization performance of the model, four types of learning models are constructed, and performance comparisons are performed for each. These models are classified as follows based on whether data augmentation techniques are applied. These are the four types of learning models: None (a model that does not apply any data augmentation and is trained based only on original A-mode data), Baseline (a model that applies only basic image augmentation techniques such as flip, blur, sharpness, and color jitter, which reflects basic image transformation methods used in general computer vision learning environments), VSPA (a model that applies only the VSPA (Variable Speed ​​and Pressure Augmentation) algorithm proposed in the present invention), and Combined (a model that applies both the Baseline augmentation technique and the VSPA algorithm, which integrates not only simple image transformation but also physics-based augmentation reflecting clinical reality).

[0086] As a result of performance evaluations conducted on an external validation set, the Combined model demonstrated the best classification performance at both the image and region levels. This is interpreted as a result of the VSPA algorithm precisely reflecting scan deviations that frequently occur in actual clinical practice, enabling the model to maintain high predictive accuracy and robustness even in new data environments where it was not trained. In particular, since VSPA is an augmentation technique closely linked to the structural characteristics of A-mode signals, it provides data diversity much closer to actual diagnostic situations than the Baseline method, which is limited to simple image distortion; this is considered to have played a key role in improving the performance of the Combined model.

[0087] Although the invention has been described with reference to the above embodiments, those skilled in the art will understand that various modifications and changes can be made to the invention without departing from the spirit and scope of the invention as described in the following claims.

[0088] [Explanation of the symbol]

[0089] 10: Probe

[0090] 20: User terminal

[0091] 100: Breast cancer detection system using an A-mode ultrasound sensor

[0092] 110: Input section

[0093] 120: Output section

[0094] 130: Communications Department

[0095] 140: Storage section

[0096] 150: Control unit

[0097] 160: Memory section

[0098] 170: Data Collection Department

[0099] 180: Data Preprocessing Section

[0100] 190: Artificial Intelligence Analysis Department

[0101] 200: User Interface Section

Claims

Input section that receives data from the outside into the system; An output unit that outputs result data to the outside; A communication unit that transmits and receives data between the inside and outside of the system; A storage unit that stores data generated by the system; A human breast cancer detection system utilizing an A-mode ultrasonic sensor, comprising: a control unit for controlling the operation of the system; and a memory unit for executing various programs and storing data therefrom. The above memory unit is a data collection unit that transmits an A-mode ultrasonic signal to the human body and receives an ultrasonic signal reflected from within the human body; A data preprocessing unit that preprocesses data collected from the above data collection unit; A human breast cancer detection system utilizing an A-mode ultrasound sensor, comprising: an artificial intelligence analysis unit that determines the presence of breast cancer based on the above-mentioned preprocessed data; and a user interface unit that provides the analysis results of the artificial intelligence analysis unit to a user. In paragraph 1, The above data collection unit is a human breast cancer detection system utilizing an A-mode ultrasonic sensor, comprising a transmission control processor that generates transmission control pulses of low voltage and low current, and an ultrasonic driving pulse amplifier that amplifies the pulses to high voltage and high current to drive an ultrasonic transducer. In paragraph 1, A human breast cancer detection system utilizing an A-mode ultrasonic sensor, wherein the data collection unit controls the switching between a transmission mode and a reception mode, and includes a transmit / receive switching switch that transmits a high-voltage pulse to a transducer during transmission and transmits a reflected signal to a receiving circuit during reception. In paragraph 1, A human breast cancer detection system utilizing an A-mode ultrasound sensor, comprising a data collection unit that performs post-processing operations including digital filtering, envelope detection, and log compression on a digitized A-mode signal. In paragraph 1, The above data preprocessing unit is a human breast cancer detection system utilizing an A-mode ultrasound sensor that performs horizontal enhancement simulating a change in scan speed to reflect user operation deviation according to the acquisition conditions of the A-mode ultrasound signal. In paragraph 5, A human breast cancer detection system utilizing an A-mode ultrasonic sensor, wherein the above horizontal direction enhancement randomly samples a resizing factor α in the range of 0.7 to 1.5, performs zero padding after compression in the horizontal direction when α is less than 1, and performs center cutting after expansion in the horizontal direction when α exceeds 1. In paragraph 1, The above data preprocessing unit is a human breast cancer detection system utilizing an A-mode ultrasonic sensor that performs vertical enhancement to reflect changes in probe pressure. In Paragraph 7, A human breast cancer detection system utilizing an A-mode ultrasonic sensor, wherein the resizing factor α is set to a value less than 1 so that the above vertical enhancement reflects only the pressure increase situation, the vertical size of the image is reduced by a factor of α, and zero padding is performed to compensate for the insufficient height. In paragraph 1, The above data preprocessing unit is a human breast cancer detection system utilizing an A-mode ultrasound sensor that includes local range compression, which performs resizing by selecting a specific horizontal area within an image to correct local signal distortion. In Paragraph 9, A human breast cancer detection system utilizing an A-mode ultrasonic sensor, wherein the above-mentioned local range compression selects an arbitrary section in the horizontal direction, applies a resizing factor α in the range of 0.1 to 0.5, and performs zero padding to the extent of the insufficient width of the compressed section. In paragraph 1, The above artificial intelligence analysis unit converts the input A-mode ultrasound signal into a two-dimensional image form and utilizes it for learning in a human breast cancer detection system using an A-mode ultrasound sensor. In paragraph 1, The above artificial intelligence analysis unit includes a ResNet50 model based on a Convolutional Neural Network, and the model is a human breast cancer detection system utilizing an A-mode ultrasound sensor that analyzes local amplitude anomalies and rapid changes in reflection intensity of the A-mode ultrasound signal. In paragraph 1, The above artificial intelligence analysis unit includes a Vision Transformer-based ViT-B / 16 model, and the model is a human breast cancer detection system utilizing an A-mode ultrasound sensor that learns the overall context and positional structure between patches of the waveform of the A-mode ultrasound signal. In paragraph 1, The above artificial intelligence analysis unit is a human breast cancer detection system utilizing an A-mode ultrasound sensor that detects structural distortions within waveforms and contextual discontinuities between signals through a Multi-head Self-Attention technique. In paragraph 1, The above artificial intelligence analysis unit is a human breast cancer detection system utilizing an A-mode ultrasound sensor based on at least one of the following criteria when determining the presence of breast cancer: abnormal amplitude interval, irregularity of reflection signal interval, sudden change in reflection intensity, and deterioration of structural consistency. In paragraph 1, The aforementioned artificial intelligence analysis unit not only handles a single A-mode image unit (image-level), but also, A human breast cancer detection system utilizing an A-mode ultrasound sensor that performs a comprehensive prediction at the region level by integrating the prediction results of multiple A-mode images collected from the same patient. In Paragraph 16, The above comprehensive prediction is a human breast cancer detection system utilizing an A-mode ultrasound sensor, which performs a soft voting method that calculates the average of the breast cancer prediction probabilities of each A-mode image. In paragraph 1, The above artificial intelligence analysis unit is a human breast cancer detection system utilizing an A-mode ultrasound sensor that performs training and validation using a five-fold cross-validation method to evaluate the generalization performance of an artificial intelligence model.

Images