Error Corrective Boosting for Learning Fully Convolutional Networks with Limited Data

Roy, Abhijit Guha; Conjeti, Sailesh; Sheet, Debdoot; Katouzian, Amin; Navab, Nassir; Wachinger, Christian

doi:10.1007/978-3-319-66179-7_27

Cited by 76 publications

(87 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To our knowledge, Roy et al [21] created the first architecture that, by combining U-Net and SegNet, segmented the whole brain slice by slice and opted for simply applying 2D convolutions. Error corrective boosting was the main key to achieve good performance but, apart from this, they did not need any additional input to get the desired results.…”

Section: Discussionmentioning

confidence: 99%

“…2b), on the other hand, train end-to-end and voxel-tovoxel in each slice. Compared to the voxel-wise CNN technique, they generate the entire segmentation from the slice input, which can better use and preserve neighborhood information in the predicted segmentation, and receives as an input the whole image [30,21,22,29]. It is also possible to extract and use patches from the whole image, resulting in the segmentation of the whole patch that is given as input.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Convolutional neural network MRI segmentation for fast and robust optimization of transcranial electrical current stimulation of the human brain

Sendra-Balcells¹,

Salvador²,

Pedro

et al. 2020

Preprint

View full text Add to dashboard Cite

The segmentation of structural MRI data is an essential step for deriving geometrical information about brain tissues. One important application is in transcranial direct current stimulation (e.g., tDCS), a non-invasive neuromodulatory technique where head modeling is required to determine the electric field (E-field) generated in the cortex to predict and optimize its effects. Here we propose a deep learning-based model (StarNEt) to automatize white matter (WM) and gray matter (GM) segmentation and compare its performance with FreeSurfer, an established tool. Since good definition of sulci and gyri in the cortical surface is an important requirement for E-field calculation, StarNEt is specifically designed to output masks at a higher resolution than that of the original input T1w-MRI. StarNEt uses a residual network as the encoder (ResNet) and a fully convolutional neural network with U-net skip connections as the decoder to segment an MRI slice by slice. Slice vertical location is provided as an extra input. The model was trained on scans from 425 patients in the open-access ADNI+IXI datasets, and using FreeSurfer segmentation as ground truth. Model performance was evaluated using the Dice Coefficient (DC) in a separate subset (N=105) of ADNI+IXI and in two extra testing sets not involved in training. In addition, FreeSurfer and StarNEt were compared to manual segmentations of the MRBrainS18 dataset, also unseen by the model. To study performance in real use cases, first, we created electrical head models derived from the FreeSurfer and StarNEt segmentations and used them for montage optimization with a common target region using a standard algorithm (Stimweaver) and second, we used StarNEt to successfully segment the brains of minimally conscious state (MCS) patients having suffered from brain trauma, a scenario where FreeSurfer typically fails. Our results indicate that StarNEt matches FreeSurfer performance on the trained tasks while reducing computation time from several hours to a few seconds, and with the potential to evolve into an effective technique even when patients present large brain abnormalities.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Convolutional neural network MRI segmentation for fast and robust optimization of transcranial electrical current stimulation of the human brain

Sendra-Balcells¹,

Salvador²,

Pedro

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…To overcome the class imbalance, a median frequency balancing is used in the class weights

w_{c} false(boldx false)

of the Dice loss to compensate for classes with low occupation probability. Furthermore, a boundary compensation introduced in is implemented in the class weights to increase the weights on the anatomical boundaries for contour correction. Thus, the class weights

w_{c} false(boldx false)

are composed of two terms: the median frequency balancing and boundary compensation:

w_{c} false(boldx false) = \frac{m e d i a n (f)}{f_{c}} + λ \cdot I false(false| ▽ G_{c} false(boldx false) false| > 0 false)

where scalar

f_{c}

is the class probability, that is, the occupation frequency of class c in the training data.…”

Section: Methodsmentioning

confidence: 99%

Automatic multi‐organ segmentation in dual‐energy CT (DECT) with dedicated 3D fully convolutional DECT networks

Chen

Zhong

et al. 2020

Medical Physics

View full text Add to dashboard Cite

Purpose Dual‐energy computed tomography (DECT) has shown great potential in many clinical applications. By incorporating the information from two different energy spectra, DECT provides higher contrast and reveals more material differences of tissues compared to conventional single‐energy CT (SECT). Recent research shows that automatic multi‐organ segmentation of DECT data can improve DECT clinical applications. However, most segmentation methods are designed for SECT, while DECT has been significantly less pronounced in research. Therefore, a novel approach is required that is able to take full advantage of the extra information provided by DECT. Methods In the scope of this work, we proposed four three‐dimensional (3D) fully convolutional neural network algorithms for the automatic segmentation of DECT data. We incorporated the extra energy information differently and embedded the fusion of information in each of the network architectures. Results Quantitative evaluation using 45 thorax/abdomen DECT datasets acquired with a clinical dual‐source CT system was investigated. The segmentation of six thoracic and abdominal organs (left and right lungs, liver, spleen, and left and right kidneys) were evaluated using a fivefold cross‐validation strategy. In all of the tests, we achieved the best average Dice coefficients of 98% for the right lung, 98% for the left lung, 96% for the liver, 92% for the spleen, 95% for the right kidney, 93% for the left kidney, respectively. The network architectures exploit dual‐energy spectra and outperform deep learning for SECT. Conclusions The results of the cross‐validation show that our methods are feasible and promising. Successful tests on special clinical cases reveal that our methods have high adaptability in the practical application.

show abstract

“…For both datasets, we trained the models with a composite equally-balanced loss function comprised of Dice loss and weighted cross entropy. The weights were labels full precision TernaryNet 3DQ computed with median frequency balancing [16] to circumvent class imbalance. We used an Adam optimizer, initialized with learning rate 0.0001 for 3D U-Net and 0.00005 for V-Net.…”

Section: Methodsmentioning

confidence: 99%

3DQ: Compact Quantized Neural Networks for Volumetric Whole Brain Segmentation

Paschali¹,

Gasperini²,

Roy

et al. 2019

Lecture Notes in Computer Science

Self Cite

View full text Add to dashboard Cite

Model architectures have been dramatically increasing in size, improving performance at the cost of resource requirements. In this paper we propose 3DQ, a ternary quantization method, applied for the first time to 3D Fully Convolutional Neural Networks (F-CNNs), enabling 16x model compression while maintaining performance on par with full precision models. We extensively evaluate 3DQ on two datasets for the challenging task of whole brain segmentation. Additionally, we showcase our method's ability to generalize on two common 3D architectures, namely 3D U-Net and V-Net. Outperforming a variety of baselines, the proposed method is capable of compressing large 3D models to a few MBytes, alleviating the storage needs in space-critical applications.

show abstract

Error Corrective Boosting for Learning Fully Convolutional Networks with Limited Data

Cited by 76 publications

References 16 publications

Convolutional neural network MRI segmentation for fast and robust optimization of transcranial electrical current stimulation of the human brain

Convolutional neural network MRI segmentation for fast and robust optimization of transcranial electrical current stimulation of the human brain

Automatic multi‐organ segmentation in dual‐energy CT (DECT) with dedicated 3D fully convolutional DECT networks

3DQ: Compact Quantized Neural Networks for Volumetric Whole Brain Segmentation

Contact Info

Product

Resources

About