Deep Conformal Supervision: Leveraging Intermediate Features for Robust Uncertainty Quantification DOI
Amir Mohammad Vahdani, Shahriar Faghani

Deleted Journal, Год журнала: 2024, Номер unknown

Опубликована: Окт. 7, 2024

Trustworthiness is crucial for artificial intelligence (AI) models in clinical settings, and a fundamental aspect of trustworthy AI uncertainty quantification (UQ). Conformal prediction as robust (UQ) framework has been receiving increasing attention valuable tool improving model trustworthiness. An area active research the method non-conformity score calculation conformal prediction. We propose deep supervision (DCS), which leverages intermediate outputs calculation, via weighted averaging based on inverse mean calibration error each stage. benchmarked our two publicly available datasets focused medical image classification: pneumonia chest radiography dataset preprocessed version 2019 RSNA Intracranial Hemorrhage dataset. Our achieved coverage errors 16e-4 (CI: 1e-4, 41e-4) 5e-4 10e-4) compared to baseline 28e-4 2e-4, 64e-4) 21e-4 8e-4, 3e-4) datasets, respectively (p < 0.001 both datasets). Based findings, results already exhibit small errors. However, shows significant improvement error, particularly noticeable scenarios involving smaller or when considering acceptable levels, are developing UQ frameworks healthcare applications.

Язык: Английский

DCN-Deeplabv3+: A Novel Road Segmentation Algorithm Based on Improved Deeplabv3+ DOI Creative Commons
Hongming Peng,

Siyu Xiang,

Mingju Chen

и другие.

IEEE Access, Год журнала: 2024, Номер 12, С. 87397 - 87406

Опубликована: Янв. 1, 2024

Язык: Английский

Процитировано

6
Medical Image Classification Using Lightweight Deep Spiking Neural Network DOI
Sandipan Bhowmick, Ashim Saha, Suman Deb

и другие.

Iranian Journal of Science and Technology Transactions of Electrical Engineering, Год журнала: 2025, Номер unknown

Опубликована: Апрель 11, 2025

Язык: Английский

Процитировано

0
Comparison of CNNs and Transformer Models in Diagnosing Bone Metastases in Bone Scans Using Grad-CAM DOI

Sehyun Pak,

Hye Joo Son, Dongwoo Kim

и другие.

Clinical Nuclear Medicine, Год журнала: 2025, Номер unknown

Опубликована: Апрель 16, 2025

Purpose: Convolutional neural networks (CNNs) have been studied for detecting bone metastases on scans; however, the application of ConvNeXt and transformer models has not yet explored. This study aims to evaluate performance various deep learning models, including in diagnosing metastatic lesions from scans. Materials Methods: We retrospectively analyzed scans patients with cancer obtained at 2 institutions: training validation sets (n=4626) were Hospital 1 test set (n=1428) was 2. The evaluated included ResNet18, Data-Efficient Image Transformer (DeiT), Vision (ViT Large 16), Swin (Swin Base), Large. Gradient-weighted class activation mapping (Grad-CAM) used visualization. Results: Both demonstrated that large model (0.969 0.885, respectively) exhibited best performance, followed by Base (0.965 0.840, respectively), both which significantly outperformed ResNet (0.892 0.725, respectively). Subgroup analyses revealed all greater diagnostic accuracy polymetastasis compared those oligometastasis. Grad-CAM visualization focused more identifying local lesions, whereas global areas such as axial skeleton pelvis. Conclusions: Compared traditional CNN superior scans, especially cases polymetastasis, suggesting its potential medical image analysis.

Язык: Английский

Процитировано

0
Deep Conformal Supervision: Leveraging Intermediate Features for Robust Uncertainty Quantification DOI
Amir Mohammad Vahdani, Shahriar Faghani

Deleted Journal, Год журнала: 2024, Номер unknown

Опубликована: Окт. 7, 2024

Trustworthiness is crucial for artificial intelligence (AI) models in clinical settings, and a fundamental aspect of trustworthy AI uncertainty quantification (UQ). Conformal prediction as robust (UQ) framework has been receiving increasing attention valuable tool improving model trustworthiness. An area active research the method non-conformity score calculation conformal prediction. We propose deep supervision (DCS), which leverages intermediate outputs calculation, via weighted averaging based on inverse mean calibration error each stage. benchmarked our two publicly available datasets focused medical image classification: pneumonia chest radiography dataset preprocessed version 2019 RSNA Intracranial Hemorrhage dataset. Our achieved coverage errors 16e-4 (CI: 1e-4, 41e-4) 5e-4 10e-4) compared to baseline 28e-4 2e-4, 64e-4) 21e-4 8e-4, 3e-4) datasets, respectively (p < 0.001 both datasets). Based findings, results already exhibit small errors. However, shows significant improvement error, particularly noticeable scenarios involving smaller or when considering acceptable levels, are developing UQ frameworks healthcare applications.

Язык: Английский

Процитировано

1