An Explainable Classification Method of SPECT Myocardial Perfusion Images in Nuclear Cardiology Using Deep Learning and Grad-CAM

Applied Sciences, Journal Year: 2022, Volume and Issue: 12(15), P. 7592 - 7592

Published: July 28, 2022

Background: This study targets the development of an explainable deep learning methodology for automatic classification coronary artery disease, utilizing SPECT MPI images. Deep is currently judged as non-transparent due to model’s complex non-linear structure, and thus, it considered a «black box», making hard gain comprehensive understanding its internal processes explain behavior. Existing artificial intelligence tools can provide insights into functionality especially convolutional neural networks, allowing transparency interpretation. Methods: seeks address identification patients’ CAD status (infarction, ischemia or normal) by developing pipeline in form handcrafted network. The proposed RGB-CNN model utilizes various pre- post-processing deploys state-of-the-art explainability tool produce more interpretable predictions decision making. dataset includes cases from 625 patients stress rest representations, comprising 127 infarction, 241 ischemic, 257 normal previously classified doctor. imaging was split 20% testing 80% training, which 15% further used validation purposes. Data augmentation employed increase generalization. efficacy well-known Grad-CAM-based color visualization approach also evaluated this research with interpretability detection infarction images, counterbalancing any lack rationale results extracted CNNs. Results: achieved 93.3% accuracy 94.58% AUC, demonstrating efficient performance stability. Grad-CAM has shown be valuable explaining CNN-based judgments nuclear physicians make fast confident using visual explanations offered. Conclusions: Prediction indicate robust based on diagnosis medicine.

Language: Английский

---

Influence Functions in Deep Learning Are Fragile

arXiv (Cornell University), Journal Year: 2020, Volume and Issue: unknown

Published: Jan. 1, 2020

Influence functions approximate the effect of training samples in test-time predictions and have a wide variety applications machine learning interpretability uncertainty estimation. A commonly-used (first-order) influence function can be implemented efficiently as post-hoc method requiring access only to gradients Hessian model. For linear models, are well-defined due convexity underlying loss generally accurate even across difficult settings where model changes fairly large such estimating group influences. functions, however, not well-understood context deep with non-convex functions. In this paper, we provide comprehensive large-scale empirical study successes failures neural network models trained on datasets Iris, MNIST, CIFAR-10 ImageNet. Through our extensive experiments, show that architecture, its depth width, well extent parameterization regularization techniques strong effects accuracy particular, find (i) estimates for shallow networks, while deeper networks often erroneous; (ii) certain architectures datasets, weight-decay is important get high-quality estimates; (iii) vary significantly depending examined test points. These results suggest general fragile call developing improved estimation methods mitigate these issues setups.

Language: Английский

---

CX-DaGAN: Domain Adaptation for Pneumonia Diagnosis on a Small Chest X-Ray Dataset

IEEE Transactions on Medical Imaging, Journal Year: 2022, Volume and Issue: 41(11), P. 3278 - 3288

Published: June 10, 2022

Recent advances in deep learning led to several algorithms for the accurate diagnosis of pneumonia from chest X-rays. However, these models require large training medical datasets, which are sparse, isolated, and generally private. Furthermore, imaging known over-fit a particular data domain source, i.e., do not conserve same accuracy when tested on dataset another center, mainly due image distribution discrepancies. In this work, adaptation classification technique is proposed overcome challenges small dataset. This method uses private-small (target domain), public-large labeled center (source consists three steps. First, it performs selection source domain's most representative images based similarity constraints through principal component analysis subspaces. Second, selected samples fit target an translation cycle-generative adversarial network. Finally, train adapted used within convolutional neural network explore different settings adjust layers perform test It shown that fine-tuning few specific together with selected-adapted increases sorting while reducing trainable parameters. The approach achieved notable increase dataset's overall accuracy, reaching up 97.78 % compared 90.03 by standard transfer learning.

Language: Английский

---

Automated Detection of Gastric Lesions in Endoscopic Images by Leveraging Attention-Based YOLOv7

IEEE Access, Journal Year: 2023, Volume and Issue: 11, P. 87166 - 87177

Published: Jan. 1, 2023

Gastric cancer is a leading cause of mortality, resulting in approximately 770000 deaths the year 2020. Early detection theatres vital role facilitating targeted treatments for gastric conditions. One commonly employed method diagnosis and treatment gastrointestinal ailments endoscopy. However, effectiveness endoscopy heavily depends on expertise endoscopist. By integrating Artificial Intelligence techniques with endoscopic procedures, we can enhance swiftness accuracy diagnostic process.This study presents an automated approach that enhances YOLO-v7 object algorithm through integration Squeeze Excitation attention block. This significantly improves small lesions, demonstrating promising results. The attention-powered YOLOv7 achieved notable precision, recall, F1-score, mean average precision values 0.72, 0.69, 0.71, respectively. Additionally, system high frame rate 63 Frames Per Second, making it well-suited real-time applications. Furthermore, performance comparison baseline model revealed 10% increase improved small-sized lesions. proposed architecture enables lesion identification, thereby supporting endoscopists analysis images, early diagnosis, reducing reliance operator's expertise.

Language: Английский

---

Attention-Based Deep Learning Approach for Breast Cancer Histopathological Image Multi-Classification

Diagnostics, Journal Year: 2024, Volume and Issue: 14(13), P. 1402 - 1402

Published: July 1, 2024

Breast cancer diagnosis from histopathology images is often time consuming and prone to human error, impacting treatment prognosis. Deep learning diagnostic methods offer the potential for improved accuracy efficiency in breast detection classification. However, they struggle with limited data subtle variations within between types. Attention mechanisms provide feature refinement capabilities that have shown promise overcoming such challenges. To this end, paper proposes Efficient Channel Spatial Network (ECSAnet), an architecture built on EfficientNetV2 augmented a convolutional block attention module (CBAM) additional fully connected layers. ECSAnet was fine-tuned using BreakHis dataset, employing Reinhard stain normalization image augmentation techniques minimize overfitting enhance generalizability. In testing, outperformed AlexNet, DenseNet121, EfficientNetV2-S, InceptionNetV3, ResNet50, VGG16 most settings, achieving accuracies of 94.2% at 40×, 92.96% 100×, 88.41% 200×, 89.42% 400× magnifications. The results highlight effectiveness CBAM improving classification importance

Language: Английский

---

Advances of deep learning in electrical impedance tomography image reconstruction

Frontiers in Bioengineering and Biotechnology, Journal Year: 2022, Volume and Issue: 10

Published: Dec. 14, 2022

Electrical impedance tomography (EIT) has been widely used in biomedical research because of its advantages real-time imaging and nature being non-invasive radiation-free. Additionally, it can reconstruct the distribution or changes electrical properties sensing area. Recently, with significant advancements use deep learning intelligent medical imaging, EIT image reconstruction based on received considerable attention. This study introduces basic principles summarizes application progress regards to three aspects: a single network reconstruction, combined traditional algorithm multiple hybrid reconstruction. In future, optimizing datasets may be main challenge applying for Adopting better structure, focusing joint algorithms, using multimodal learning-based solution existing problems. general, offers fresh approach improving performance could foundation building an integrated diagnostic system future.

Language: Английский

---

Healthcare transformed: a comprehensive survey of artificial intelligence trends in healthcare industries

Elsevier eBooks, Journal Year: 2024, Volume and Issue: unknown, P. 395 - 424

Published: Jan. 1, 2024

Language: Английский

---

Polycystic Ovary Syndrome (PCOS) diagnostic methods in machine learning: a systematic literature review

Multimedia Tools and Applications, Journal Year: 2024, Volume and Issue: unknown

Published: June 27, 2024

Language: Английский

---

Local interpretable model-agnostic explanation approach for medical imaging analysis: A systematic literature review

Computers in Biology and Medicine, Journal Year: 2024, Volume and Issue: 185, P. 109569 - 109569

Published: Dec. 19, 2024

Language: Английский

---

The Present and Future of Artificial Intelligence-Based Medical Image in Diabetes Mellitus: Focus on Analytical Methods and Limitations of Clinical Use

Journal of Korean Medical Science, Journal Year: 2023, Volume and Issue: 38(31)

Published: Jan. 1, 2023

Artificial intelligence (AI)-based diagnostic technology using medical images can be used to increase examination accessibility and support clinical decision-making for screening diagnosis. To determine a machine learning algorithm diabetes complications, literature review of studies image-based AI was conducted the National Library Medicine PubMed, Excerpta Medica databases. Lists as keywords were combined. In total, 227 appropriate selected. Diabetic retinopathy model most frequent (85.0%, 193/227 cases), followed by diabetic foot (7.9%, 18/227 cases) neuropathy (2.7%, 6/227 cases). The open datasets (42.3%, 96/227 or directly constructed data from fundoscopy optical coherence tomography (57.7%, 131/227 Major limitations in AI-based detection complications lack (36.1%, 82/227 severity misclassification (26.4%, 60/227 Although it remains difficult use fully trust imaging analysis clinically, reduces clinicians' time labor, expectations its decision-support roles are high. Various collection synthesis developments according disease required solve imbalance.

Language: Английский

---