A Spatio-Temporal Fully Convolutional Network for Breast Lesion Segmentation in DCE-MRI

Chen, Mingjian; Zheng, Hao; Lu, Changsheng; Tu, Enmei; Yang, Jie; Kasabov, Nikola

doi:10.1007/978-3-030-04239-4_32

Cited by 13 publications

(14 citation statements)

References 19 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The architecture of deep convolutional networks was optimized using a large training dataset of over 38,000 examinations, which consisted of 2,555 malignant and 60,108 benign breast scans. This dataset is substantially larger than previous efforts with deep networks, which have used only 50-250 MRI examinations (19)(20)(21)(22)(23)(24).…”

Section: Discussionmentioning

confidence: 99%

“…Our final U-Net implementation differs from previous approaches in two important ways. First, we used a full 3D network instead of a conventional 2D network that processes individual sections independently (12,21,(23)(24)(25). While this increased the number of network parameters, it also captured volumetric features missed in 2D processing.…”

Section: Discussionmentioning

confidence: 99%

“…Additionally, background parenchymal enhancement (BPE) measured in MRI with dynamic contrast enhancement (DCE) is predictive of cancer risk (18). Several studies have automated tumor segmentation in breast MRI using modern deep networks such as U-Nets or DeepMedic (19)(20)(21)(22)(23)(24)(25), focusing mostly on malignant tumors. With reliable, fully automated segmentation, the overall clinical workflow could be improved, and such segmentation could aid radiologists in tumor detection and diagnosis.…”

Section: Author Contributionsmentioning

confidence: 99%

“…Fast automated segmentation might help identify important prognostic and predictive biomarkers. Unfortunately, the available MRI datasets to train segmentation algorithms have remained comparatively small with 50-250 MRI examinations (19)(20)(21)(22)(23)(24)(25), thus limiting the potential of modern deep networks. Some studies have been limited to semiautomated segmentation (26) and performance differs across datasets making comparison to radiologists difficult.…”

Section: Author Contributionsmentioning

confidence: 99%

“…We used networks based on the DeepMedic network (27) and a 3D U-Net (20), which have been used extensively for medical segmentation, including breast segmentation (19)(20)(21)(22)(23)(24). The architecture of the 3D U-Net is described in Figure 3 Following previous studies (28) the traditional space-invariant implementation has been augmented by adding a spatial prior as input to the final classification.…”

Section: Convolutional Neural Networkmentioning

confidence: 99%

See 4 more Smart Citations

Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans

Hirsch¹,

Huang²,

Luo³

et al. 2022

Radiology: Artificial Intelligence

View full text Add to dashboard Cite

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Discussionmentioning

confidence: 99%

Section: Author Contributionsmentioning

confidence: 99%

Section: Author Contributionsmentioning

confidence: 99%

Section: Convolutional Neural Networkmentioning

confidence: 99%

See 3 more Smart Citations

Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans

Hirsch¹,

Huang²,

Luo³

et al. 2022

Radiology: Artificial Intelligence

View full text Add to dashboard Cite

show abstract

Multiparametric deep learning tissue signatures for a radiological biomarker of breast cancer: Preliminary results

et al. 2019

View full text Add to dashboard Cite

Advances in knowledge:1. Advanced computational methods using multiparametric deep learning (MPDL) with multiparametric MRI are significant predictors of malignant or benign breast lesions.2. Development of a "tissue based" deep learning method validated with an independent radiological data set.3. The MPDL model with pharmacokinetic modeling parameters and diffusion weighted imaging/Apparent Diffusion Coefficient metrics demonstrated similar diagnostic performance of the radiologist in characterizing breast lesions.Implications for patient care: The integration of advanced computational techniques and artificial intelligence methods to assist radiologists will become available in the future reading rooms and will transform medicine in general. Deep learning methods will be the conduit for modeling of clinical and radiological variables which will provide the foundation for radiological precision medicine in patients. AbstractBackground and Purpose: Integration of advanced computational techniques and artificial intelligence methods to assist radiologists will become available in future reading rooms and will transform medicine in general. This study was conducted to evaluate the feasibility and role of a novel deep learning method using multiparametric breast Magnetic Resonance Imaging(mpMRI) and defining tissue signatures for improved automated detection and characterization of breast lesions. Methods: We developed and tested a multiparametric deep learning(MPDL) network for segmentation and classification of breast MRI in 171 patients. The MPDL network was constructed from stacked sparse autoencoders. MPDL network inputs were T1 and T2-weighted imaging, diffusion weighted imaging(DWI) and ADC mapping, and dynamic contrast enhanced(DCE) imaging tissue signatures. Evaluation of MPDL consisted of cross-validation, sensitivity and specificity. Dice similarity between MPDL and post-DCE lesions were evaluated. Statistical significance was set at p≤0.05. Results: The performance of MPDL on Modified National Institute of Standards and Technology database(MISNT) data set was 99.7%. For MRI validation set, a 4.2%±3.6% percent difference between volumes was found between MPDL and test data set. The MPDL segmented glandular, fatty, and lesion tissue with an overlap of 0.87±0.05 for malignant patients and 0.85±0.07 for benign patients. The sensitivity and specificity for differentiation of malignant from benign lesions were 90% and 85% respectively with an AUC of 0.93 Conclusion: Integrated MPDL method accurately segmented and classified different breast tissue from multiparametric breast MRI. Deep learning can be used to construct a personalized database of tissue signatures with accurate characterization of different tissue types.

show abstract

Deep Learning Based Segmentation of Breast Lesions in DCE-MRI

Khaled

Vidal

Martí

2021

Pattern Recognition. ICPR International Workshops and Challenges

View full text Add to dashboard Cite

A Spatio-Temporal Fully Convolutional Network for Breast Lesion Segmentation in DCE-MRI

Cited by 13 publications

References 19 publications

Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans

Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans

Multiparametric deep learning tissue signatures for a radiological biomarker of breast cancer: Preliminary results

Deep Learning Based Segmentation of Breast Lesions in DCE-MRI

Contact Info

Product

Resources

About