Paper Submission & Registration
8th Dutch Bio-Medical Engineering Conference
Go-previous
16:30   Lung & Respiration - I
Chair: Peter Veltink
16:30
15 mins
Multi-planar deep convolutional networks for nodule detection in CT scans
Sunyi Zheng, Ludo J. Cornelissen, Xiaonan Cui, Xueping Jing, Raymond N. J. Veldhuis, Matthijs Oudkerk, Peter M.A. van Ooijen
Abstract: Purpose: Lung cancer is one of the most malignant cancers. Early detection of lung nodules can give better treatment alternatives to patients and increase their survival chances. In clinical evaluation, radiologists would commonly use the axial, coronal and sagittal planes, rather than solely the axial plane to detect nodules. Our study aimed to explore whether a multi-planar approach can increase the performance of a deep learning-based system for automated nodule detection and thus provide better assistance for radiologists. Materials and Methods: We proposed a deep learning-based detection (DL-CAD) system for pulmonary nodule detection. The system comprised two stages, nodule candidate detection and false positive reduction. At the first stage, a deeply-supervised encoder-decoder network, U-net plus plus, was applied to find possible nodule candidates using 1 mm axial, coronal, and sagittal slices and 10 mm axial maximum intensity projection slices. To eliminate false positive findings a 3-D multi-scale dense convolutional neural network that extracted multi-scale contextual information was then designed. The method was trained and evaluated based on the public LUNA16 dataset. The dataset contains 888 CT scans with 1186 nodules collected from seven academic centers. Four radiologists annotated the dataset independently. The reference standard comprises those nodules (>3 mm) that were accepted by at least three out of four radiologists. The performance of our proposed system was evaluated using ten-fold cross-validation. Results: Our multi-planar approach achieved a sensitivity of 96.0% with 2 false positive per scan on the public LUNA16 dataset. The designed DL-CAD system detected 98.1% nodules when results were merged from multiple planes, which outperforms the sensitivity from any single plane (1 mm axial slices: 91.1%; 1 mm coronal slices: 82.5%; 1 mm sagittal slices: 81.8%; 10 mm MIP slices: 93.2%). Although small nodules (i.e. <6 mm) are notoriously hard to detect, the proposed CAD system detected 97% of these nodules. Conclusion: The multi-planar method substantially improved the number of detected nodules compared to that using only the axial plane. The results showed the effectiveness of the system and its potential to assist radiologists in clinical practice.
16:45
15 mins
Detection and classification of patient-ventilator asynchrony
Tom Bakkes, Roel Montree, Massimo Mischi, Francesco Mojoli, Simona Turco
Abstract: Patient-ventilation asynchrony (PVA) is a mismatch that can occur during mechanical ventilation (MV), when patient inspiratory and expiratory efforts are not in synchrony with the MV cycle. These mismatches are the result of incorrect detection of the patients’ respiratory efforts, resulting in the ventilator supplying air pressure and flow at the wrong intervals. This can cause higher or lower air pressures and/or volumes in the patients’ lungs, which in turn can cause discomfort, barotrauma, and volumetric distention [1]. Therefore, proper detection of these asynchronies is needed to decrease the risks associated with MV. In this abstract the study performed in [2] is summarised. In this study we focussed on detecting and classifying PVA. The method works in three steps. In the first step a neural network with the u-net architecture [3] was utilized to detect the start and end of the patients’ respiratory efforts from the air pressure, flow, and volume waveforms. The network was trained and tested using cross-validation on 15 patients containing a total of 4275 breaths. In the second step the start and end of these patient respiratory efforts were compared to the start and end of the MV. In the third step the timing differences between the patient and MV were utilized to classify the type of asynchrony. In our research the distinction was made between the following asynchronies: delayed inspiration, early cycling, late cycling, and ineffective efforts. The detection method was able to achieve a sensitivity and precision of 98.6% and 97.3%, respectively, for the start of the patients’ respiration, and 97.7% and 97.2%, respectively, for the end of the patients’ respiration. The classification had mixed results based on the type of asynchrony with F1-scores ranging from 67.8% to 95.0%. In the future we hope to able to utilize these techniques to aid clinicians in setting up and adjusting ventilator parameters to reduce the number of asynchronies during MV. [1] P. Spieth et. al., Dtsch Artztebl Int. (2014) [2] T. Bakkes et. al., EMBC. (2020) [3] O. Ronneberger et. al. LNCS (2015)
17:00
15 mins
Recurrent convolutional neural network for snore detection using audio data
Jiali Xie, Xavier Aubert, Long Xi, Johannes van Dijk, Bruno Arsenali, Pedro Fonseca, Sebastiaan Overeem
Abstract: Research question: Snoring is a prevalent phenomenon affecting about 30% of the adult population. It may be benign, but can also be a symptom related to obstructive sleep apnea (OSA), a prevalent sleep disorder. Accurate detection of snoring may help with screening and diagnosis of OSA. In this study, we introduce a robust snore detection method based on audio recordings. Methods: The snore detection algorithm is based on the combination of a convolutional neural network (CNN) and a recurrent neural network (RNN). We obtained audio recordings of 38 subjects referred to a clinical center for sleep study. All subjects underwent a full PSG at Kempenhaeghe Center for Sleep Medicine in the Netherlands. Subjects were recorded by a total of 5 microphones placed at strategic positions around the bed. The CNN was used to extract features from the sound spectrogram, while the RNN was used to process the sequential CNN output and to classify the audio events to snore and non-snore events. Acoustic analysis was performed along a non-linear frequency scale providing a detailed resolution of the lower part of the audio spectrum. Results: The algorithm achieved an accuracy of 95.3 ± 0.5%, a sensitivity of 92.2 ± 0.9%, and a specificity of 97.7 ± 0.4% over all microphones in snore detection on our data set including 18412 sound events. The best accuracy (95.9%) was observed from the microphone placed about 70 cm above the subject’s head and the worst (94.4%) was observed from the microphone placed about 130 cm above the subject’s head. Conclusions: Our results suggest that our method detects snore events from audio recordings with high accuracy and is robust with audio data from all 5 microphones.
17:15
15 mins
Prognostic outcome prediction for early stage non-small cell lung cancer using deep learning
Sunyi Zheng, Jiapan Guo, Johannes A. Langendijk, Stefan Both, Ludo J. Cornelissen, Raymond N. J. Veldhuis, Matthijs Oudkerk, Peter M.A. van Ooijen, Robin Wijsman, Nanna M. Sijtsema
Abstract: Research question: Standard treatment of early stage non-small cell lung cancer (NSCLC) is lobectomy. However, Stereotactic Body Radiation Therapy (SBRT) is a good alternative for patients that are not fit enough for surgery with 3-year recurrence rates (local, regional, and distant recurrence) between 18 and 29%. The goal of this study was to develop a deep learning model to provide prognostic predictions by extracting image features from pre-treatment CT-scans for early stage NCSLS patients treated with SBRT . Such a model could be used to identify patients with worse outcomes to be able to optimize their treatment in the future. Materials and Method: Training and validation cohorts, including 164 and 119 patients who received SBRT, between February 2013 and August 2018 at the University Medical Center Groningen were prospectively collected. The endpoints used for analysis were local recurrence (LR), regional recurrence (RR), distant metastasis (DM), overall survival (OS), tumor-specific survival (TSS) and disease-free survival (DFS). The 3D deep learning model for outcome predictions at 2 years was implemented based on the Resnet-18. Cubic patches (64×64×64) containing lung tumors extracted from pre-treatment CT scans were used for training. The performance of the model was assessed using the receiver operating characteristic curve. Sensitivity, specificity, the positive predictive value (PPV), and the negative predictive value (NPV) were determined for cut-off values resulting in the maximum sum of sensitivity and specificity for the endpoint. Results: Table 1 shows the performance of the deep learning model on LR, RR, DM, OS, TSS and DFS. The area under the curve for LR prediction was 0.74. For the RR, DM, and DFS prediction, the sensitivity was above 84%, while the specificity ranged from 30% to 40%. The model achieved specificities larger than 61% on LR, OS, TSS, whereas the sensitivities were between 47% and 75%. Moreover, the NPV of the model ranged from 78% to 97% and the PPV was between 8% and 47%. Conclusions: The preliminary results showed the deep learning model could identify prognostic image features from pre-treatment CT-images. The performance of the LR prediction was promising.


end %-->