Research

Impact of computed tomography reconstruction kernel selection on the performance of deep learning-based lung nodule detection

In collaboration with Paras Hospitals, Gurugram, New Delhi, India; 

ECRTo be Presented: 11-15 March 2020, European Congress of Radiology 2020

Oral Presentation

Purpose

To evaluate the relationship between computed tomography reconstruction kernels and the sensitivity of deep learning algorithms for lung nodule detection. We attempted to understand consistencies and discrepancies for the same.

Methods

A 20-layer deep residual convolutional neural network based on 3D Feature pyramid networks was trained and validated on 888 CTs from the LIDC-IDRI dataset for lung nodule detection. The LIDC-IDRI dataset had a good mixture of CTs with different reconstruction kernels: soft-tissue (58%), lung (27%), and sharp (13%). 100 pairs of CTs with soft-tissue and lung kernels were randomly sampled from the NLST dataset. Ground truth for the same studies was annotated by a pair of junior and senior radiologists, with the AI’s prediction used as priors. A post-study analysis was done to measure the relationship between sensitivity and the reconstruction kernels (soft-tissue and lung) used in the CTs from NLST at different false-positive rates.

Results

The AI showed a sensitivity of 88% at 1 FP/scan for detection of >= 4 mm nodules on 177 LIDC-IDRI CTs when compared against the consensus of 3 out of 4 radiologists, and a sensitivity of 75% at 1.5 FPs/scan for 100 lung kernel CTs and 75% at 1.6 FPs/scan for 100 soft-tissue kernel CTs from NLST with AI + radiologists as ground truth. The difference in performance by the AI in soft-tissue kernel CTs and lung kernel CTs was not found to be statistically significant as the one-way ANOVA revealed p=0.9938 and p=0.90, respectively, for sensitivity and false-positives per scan.

Conclusion

Deep learning algorithms are proving to be robust in the detection of lung nodules between CTs reconstructed from different kernels.

A comparison of lung nodule detection sensitivity of deep learning algorithms in comparison with 3 radiologists of varying experience levels

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRTo be Presented: 11-15 March 2020, European Congress of Radiology 2020

Oral Presentation

Purpose

To compare the performance of deep learning algorithms for lung nodule detection against 3 independent radiologists of different experience levels (28 yrs, 20 yrs, and 14 yrs) on a retrospective chest CT dataset of 240 patients

Methods

In a retrospective clinical validation study, three independent radiologists were required to annotate 240 chest CTs for all visible nodules between 5-30 mm in size. The annotations were compared with outputs of a deep learning algorithm (Predible Health) for nodule detection All nodules were sub-classified as having consensus among 1, 2, or 3 radiologists. The deep learning algorithm was individually tested on nodule datasets with varying consensus levels. When testing on the 2/3 and 3/3 consensus dataset, we marked the 1/3 consensus nodules as irrelevant findings and did not penalise the algorithm for its detection.

Results

After considering for overlaps, there were a total of 146 nodules marked by the radiologists. The algorithm had a sensitivity of 96% (47/49) on nodules with 3/3 consensus and sensitivity of 91.1% (72/79) on nodules with 2/3 consensus, both at 1 false-positives per scan. The algorithm had an overall FROC score of 0.91 on nodules with 3/3 consensus and 0.86 on nodules with 2/3 consensus. Combining the algorithms’ findings with each radiologist' and comparing with the consensus of the other two as ground truth, we observe an increased sensitivity ranging from 5 to 20%.

Conclusion

This study demonstrates that lung nodule detection algorithms can improve the sensitivity of radiologists by as much as 20%, helping them report quicker.

Takeaways from the validation of an AI-based malignancy likelihood estimation for lung cancer screening when used on routine CT studies in a tertiary care hospital

In collaboration with Max Super Speciality Hospital, Saket, New Delhi, India; 

ECRTo be Presented: 11-15 March 2020, European Congress of Radiology 2020

Oral Presentation

Purpose

To assess the performance of a deep learning-based malignancy likelihood estimation system trained on screening data from the National Lung Cancer Screening Trial (NLST) on a retrospective dataset of biopsy-proven studies from a large tertiary hospital in North India. The retrospective clinical validation dataset consisted of cases with primary cancers (64%), metastatic cancers (13%), and benign conditions (23%).

Methods

The deep learning algorithm was trained on 1,245 scans from the NLST trial with pathologically proven ground truths to determine the malignancy status of a lung nodule. Retrospective data of 123 patients over 20 months who underwent CT-guided lung biopsy were chosen as suitable for the validation study. All patient studies were evaluated on follow-up scans to ensure conformance with biopsy results.

Results

The AI model showed a sensitivity of 75% (95% CI: 64%-83%) on 95 malignant nodules with a PPV of 86% (95% CI: 79%-90%). It is notable that 6 studies that appeared negative on histological findings were correctly predicted by the AI model as malignant, as confirmed by follow-up scans and re-biopsy. The AI model had a specificity of 58% (95% CI: 37%-76%) on 28 benign nodules with an NPV of 40% (95% CI: 29%-51%).

Conclusion

The algorithm can be positioned as an adjunct tool to the histology report that sometimes suffers an error rate due to erroneous sampling. Cancer screening settings only account for a limited context in training data due to specific inclusion criteria and hence further enhancement using a larger and more diverse training dataset is required before it can be used in routine practice.

Establishing Normative Liver Volume and Attenuation for the Population of a Large Developing Country Using Deep Learning – A Retrospective Study of 1,800 CT Scans

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRTo be Presented: 11-15 March 2020, European Congress of Radiology 2020

Poster

Purpose

Deep learning has enabled the analysis of large datasets which previously required significant manual labor. We used a deep learning algorithm to study distribution of liver volumes and attenuation of a massive dataset of ~1,800 non-contrast CTs (NCCTs) of the abdomen.

Methods

1,823 NCCTs of the abdomen with no liver or related abnormality on clinical reports were retrospectively pulled from PACS. Liver Volume (LV) and Mean Liver Attenuation (MLA) were determined on these scans using a deep learning algorithm which automatically segments liver tissue and subsequently calculates LV and MLA. The algorithm was tested independently on 200 CTs from LiTS challenge and gave a DICE score of 95%, and mean volume error of 3.8%. Appropriate statistical analysis (correlations, histogram etc.) were performed and estimated prevalence of fatty liver was calculated using a cut-off of 40HU, as described in literature.

Results

107 (6%) NCCTs failed the algorithm’s quality check and were excluded from the study. The remaining 1,715 NCCTs gave a mean LV of 1,389mL (SD: 473, Range: 201 – 3946) and MLA of 59.2HU (SD: 15.9, Range: 24.2 – 125.6). 122 of 1715 (females: 41%) had fatty liver (MLA < 40). There is no strong correlation between volume and age for both men (R2: 0.002) and women (R2: 0.0001). Further analysis to correlate height and weight of patients is currently underway and shall be presented at the conference. Our prevalence results closely match reported literature for the country.

Conclusion

Automated analysis using deep learning algorithms can help parse through massive datasets automatically and shed light on important clinical questions such as establishment of age- and sex-correlated normative values. We establish new normative values for LV and MLA, and quantify the prevalence of fatty liver.

Unboxing AI - Radiological Insights Into a Deep Neural Network for Lung Nodule Characterization

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

Academic RadiologyPublished: 27 January 2020, Academic Radiology

Journal

Rationale and Objectives

To explain predictions of a deep residual convolutional network for characterization of lung nodule by analyzing heat maps.

Materials and Methods

A 20-layer deep residual CNN was trained on 1245 Chest CTs from National Lung Screening Trial (NLST) trial to predict the malignancy risk of a nodule. We used occlusion to systematically block regions of a nodule and map drops in malignancy risk score to generate clinical attribution heatmaps on 103 nodules from Lung Image Database Consortium image collection and Image Database Resource Initiative (LIDC-IDRI) dataset, which were analyzed by a thoracic radiologist. The features were described as heat inside nodule -bright areas inside nodule, peripheral heat continuous/interrupted bright areas along nodule contours, heat in adjacent plane -brightness in scan planes juxtaposed with the nodule, satellite heat - a smaller bright spot in proximity to nodule in the same scan plane, heat map larger than nodule bright areas corresponding to the shape of the nodule seen outside the nodule margins and heat in calcification.

Results

These six features were assigned binary values. This feature vector was fedinto a standard J48 decision tree with 10-fold cross-validation, which gave an 85 % weighted classification accuracy with a 77.8% True Positive (TP) rate, 8% False Positive (FP) rate for benign cases and 91.8% TP and 22.2% FP rates for malignant cases. Heat Inside nodule was more frequently observed in nodules classified as malignant whereas peripheral heat, heat in adjacent plane, and satellite heat were more commonly seen in nodules classified as benign.

Conclusion

We discuss the potential ability of a radiologist to visually parse the deep learning algorithm generated "heat map" to identify features aiding classification.

Read full paper

Evaluating Appropriate Role of Artificial Intelligence in Preoperative Abdomen CT Assessment for Living Donor Liver Transplants (LDLT)

In collaboration with Max Super Speciality Hospital, Saket, New Delhi, India; 

RSNAPresented: 2 December 2019, RSNA 2019

Oral Presentation

Purpose

In LDLT, assuring appropriate graft size via evaluation of liver and segmental volumes is a major predictor of safe, successful outcomes. The analysis comprises of two key steps: 1. Segmentation of liver and hepatic vascular structures, and 2. Liver Resection to calculate graft and remnant volumes. Here we aim to study preoperative LDLT assessment using 3 different approaches: A: Fully Manual (Hepatic anatomy is segmented by manual contouring followed by manual resection), B: AI with Manual Resection (Hepatic anatomy is automatically segmented using AI and a radiologist resects manually), and C: Fully Automated (Hepatic anatomy is automatically segmented and resected by AI with no radiologist intervention).

Methods and Materials

Our developed AI system comprised of 3 CNN models trained on 324 triphasic contrast-enhanced CTs and validated on 100 CTs from multiple institutions for liver and veins segmentation and middle hepatic vein (MHV) classification. For automated resection (C), we sample points from the MHV and IVC to draw a resection plane and return the graft and remnant volumes. 100 retrospective abdomen CT scans with preoperative analysis done were extracted from a large tertiary hospital. 6 studies were excluded due to incomplete information. On the remaining 94 CTs, the graft and remnant volumes were generated for A, B, and C. Intraoperative surgical weights were collected for comparison as ground truth.

Results

We measured the variance of graft volume for A, B, and C against intraoperative surgical weight. B has the least overall variance of 9.14%, followed by C (9.32%) and A (10.62%) on 94 cases. A close correlation (variance < 5%) with the weight was seen in 40 cases using C as compared to 39 cases using B and 32 cases using A. Fig 1 shows the boxplot of the variance of A, B, and C.

Conclusion

Amongst the 3 approaches for LDLT analysis, AI with Manual Resection (B) and Fully Automated (C) give the best results, with B displaying the least overall variance.

Clinical Relevance

While AI can automate routine mundane tasks such as hepatic structure segmentation, an AI system coupled with expert intervention is poised to deliver better outcomes in Liver Transplant Planning.

Visit event page

The Subtle Art of Accurate Natural Language Processing for Radiology Report Mining

RSNAPresented: 2 December 2019, RSNA 2019

Certificate of Merit Poster

Teaching Points

With increasing usage of artificial intelligence algorithms, extraction of relevant radiology imaging becomes critical for bothdevelopment and clinical validation processes. In this exhibit, we aim to cover the following teaching points: 1) Understanding andanalysis of your radiology report dump, 2) How to write an accurate text parser with appropriate negations, and 3) How to writeyour own rules to build your text parsing engine

Outline

Introduction to NLP and Python, Text reading and parsing in python, Understanding the semantics of radiology reports, Introductionto N-gram analysis, Using n-gram analysis to find relevant information from reports (negations, tags, etc), Designing the NLP,Coding and execution on a sample dataset, Validation of NLP rules, Ensuring edge cases

Read full presentation

How to Train Your 3D U-Net for Organ Segmentation

RSNAPresented: 2 December 2019, RSNA 2019

Certificate of Merit Poster

Teaching Points

CT and MRI scans are full of 3D images. Annotation of such data with pixel-wise segmentation labels causes difficulties, since only2D slices can be shown on a computer screen. Thus, annotation of large volumes in a slice-by-slice manner is very tedious. In thisexhibit, we will delve into the intricacies of training a deep network that learns to generate dense volumetric segmentations. Weaim to cover the following teaching points: 1. How to pre-process your 3D images in Python, 2. Training the 3D U-Net with the pre-processed images, and 3. Generating segmentations on unseen CTs with the trained 3D U-Net

Outline

Introduction to CTs and MRIs as 3D volumes Reading and pre-processing the images with Python Writing a data loader in PyTorchfor 3D patches Defining the training pipeline with 3D U-Net Training the 3D U-Net and saving the checkpoints Predicting on anunseen image with the trained 3D U-Net

Read full presentation

Deep Learning for Lung Cancer Detection

In collaboration with Intel, India; 

Intel AIPublished: 20 September 2019, Intel AI

Whitepaper

With an annual incidence of 3.34 million, lung cancer is the deadliest cancer, with an estimated 1.88 million deaths per year worldwide. Early detection is critical towards long-term survival; stage 4 lung cancer has a 5-year survival rate of 5% in comparison Screening Trial (NLST) revealed that participants who received low-dose helical CT (computed tomography) scans had a 20 percent lower risk of dying from lung cancer than participants who received standard chest X-rays. Advances in multi-detector CT scanning have made high-resolution volumetric imaging possible in a single breath hold, at acceptable levels of radiation exposure. Several observational studies have shown that a low-dose helical CT scan of the lung detects more nodules and lung cancers, including early-stage cancers. Potentially malignant lung nodules can be identified from chest CT scans, and early intervention can result in a higher chance of long-term survival.

A typical chest CT scan contains anywhere in the range of 300-500 slices, and a radiologist must examine each slice to detect lung nodules. Lung nodules are small masses of tissue in the lung that appear as round, white shadows on a CT scan and are often difficult to detect and document. Most are benign, but their detection requires specialized expertise and with widespread implementation of lung cancer screening programs, the burden on radiologists is rapidly increasing.

Computer-aided-detection (CAD) is becoming increasingly useful in helping radiologists interpret high-dimensional imaging data like CT and MRI scans. CAD algorithms have also shown to be successful in increasing radiologists’ ability to detect lung nodules. With the advent of deep learning and convolutional neural networks (CNNs), CAD algorithms have started moving away from a reliance on hand-crafted features requiring custom engineering, to learning features from data through CNNs.

This whitepaper will detail how Predible Health’s deep learning algorithm for detecting lung nodules on CT scans has been optimized on Intel® Xeon® Scalable processors using the Intel® Distribution of OpenVINO™ Toolkit.

Read full paper

Opening the "Black Box": radiological insights into a deep neural network for lung nodule

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRPresented: 1 March 2019, ECR 2019

Oral Presentation

Purpose

To explain predictions of a deep residual convolutional network for characterization of lung nodule by analysing heat maps.

Methods and Materials

A 20-layer deep residual CNN was trained on 1245 chest CTs from NLST trial to predict the malignancy risk of a nodule. We used occlusion to systematically block regions of a nodule and map drops in malignancy risk score to generate clinical attribution heatmaps on 160 nodules from LIDC-IDRI dataset, which were analysed by a thoracic radiologist. The features were described as heat inside nodule (IH) - bright areas inside nodule, peripheral heat (PH) - continuous/interrupted bright areas along nodule contours, heat in adjacent plane(AH) - brightness in scan planes juxtaposed with the nodule, satellite heat (SH) - a smaller bright spot in proximity to nodule in the same scan plane, heat map larger than nodule (LH) - bright areas corresponding to the shape of the nodule seen outside the nodule margins and heat in calcification (CH).

Results

These six features were assigned binary values. This feature vector was fed into a standard J48 decision tree with tenfold cross-validation, which gave an 85% weighted classification accuracy with a 77.8%TP rate, 8% FP rate for benign cases and 91.8% TP and 22.2%FP rates for malignant cases. IH was more frequently observed in nodules classified as malignant whereas PH, AH, and SH were more commonly seen in nodules classified as benign.

Conclusion

We discuss the potential ability of a radiologist to visually parse the deep learning algorithm-generated 'heat map' to identify features aiding classification.

Visit event page

Cloud-based semi-automated liver segmentation: analytical study to compare its speed and accuracy with a semi-automated workstation based software

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRPresented: 1-3 March 2019, ECR 2019

poster

Purpose

The purpose of this study is to evaluate the performance of a fully automated post processing solution, based on deep neural networks, for liver on MDCT image datasets. We compare time taken to perform liver volumetry between (1) Manual segmentation using commercially available software and (2) Automated segmentation with manual refinement.

Methods and Materials

A test dataset of 15 multi-phasic contrast-enhanced CT scans was provided by Centre for Advanced Research in Imaging, Neurosciences and Genomics (CARING), New Delhi,India.All studies were acquired on a 128-MDCT GE Discovery IQ scanner. The images were acquired using a matrix size of 512 x 512 pixels, at an in-plane pixel size of 0.76 mm, reconstructing 0.6 mm thin images. Individual contrast bolus-tracking was performed during repetitive low dose acquisitions at 120 kVp /40 mAs and placement of a threshold region-of-interest (ROI) within the abdominal aorta at the level of the diaphragm, plotting HU contrast wash-in to a level of 150 HU following contrast administration of 100 ml 320 mg I/ml contrast agent administered at 4 ml / sec injected into a right antecubital vein using a CTA injector. The diagnostic arterial and portal-venous cranio-caudal helical hepatic MDCT acquisition commenced 12 seconds and 60 seconds post 150 HU washin, respectively.We performed liver volumetry on two setups (1) A commercially available CT Volume Viewer Package and (2) PredibleLiver (Predible Health, Bengaluru, India), a livervolumetry software package with segmentations initialized using DNNs (Fig. 2). All quantitative volumetric evaluations were performed by a radiologist (MD) of 7 years of experience. The radiologist performed manual and automated volumetry with an interval of 2 months.

Results

We can see the volumes (in ml) obtained for the two setups. The liver volumes for the studies are consistent over the two setups with a maximum variation of 2.3% and an average variation of 0.9%. We can also see the time taken (in mins) for performing liver volumetry on the two setups. Automation of liver volumetry accelerated the post processing significantly. Automated Liver volumetry on PredibleLiver takes an average of 3.5 minutes compared to 14.6 minutes on Commercial CT Volume Viewer.

Conclusion

The study shows that liver volumetry post processing can be significantly accelerated by initializing with Deep Learning based segmentation. We compared two setups : (1) Commercial CT Volume Viewer and (2) PredibleLiver. We found PredibleLiver to require lesser time in performing volumetric assessment over 15 studies as the segmentations come pre-initalized using Deep Learning.

Read full poster

Towards radiologist-level malignancy detection on chest CT scans: a comparative study of the performance of convolutional neural networks and four thoracic radiologists

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRPresented: 1-3 March 2019, ECR 2019

poster

Purpose

The purpose of this work is to evaluate the performance of a deep learning system based on convolutional neural networks in predicting the presence of malignant lung nodules on chest CT scans. We also attempt to benchmark its performance against four radiologists.

Methods and Materials

100 unseen low-dose CT scans from the validation set were chosen at random and predictions were generated from the deep learning system. Studies were randomized and presented to 4 thoracic radiologists with 2, 5, 8- and 15-years' experience to characterize the chest CT scans. The radiologists were asked to assess the probability of malignancy in the scans on a Likert scale of 1 (highly unlikely) to 5 (highly suspicious). The ROC curves were analysed for the AI and the radiologists. Post-analysis, 4 CT scans without lung nodules but marked malignant in the NLST EMR were removed from the study.

Results

The results of radiologists on the Likert scales 1 and 2 were considered as negative for malignancy and 3,4 and 5 were considered to be positive for the presence of malignancy. For the AI, a predicted probability > 0.25 was considered to be positive for the presence of malignancy.

On the 96 chest CT scans reviewed by the radiologists, they had AUCs of 0.82, 0.81, 0.83 and 0.83 for predicting the risk of malignancy whereas the AI had an AUC of 0.91. Individually, radiologists' accuracy varied from 76 to 77% and AI's accuracy was 83%. The difference in the radiologist's interpretation was not found to be statistically significant as the one - way ANOVA revealed p-value is 0.77311.

Conclusion

The deep learning system shows better performance than experienced radiologists, individually and in aggregate, in predicting the presence of malignant nodules on the 96 CT scans obtained from the NLST dataset. The difference in the interpretation of radiologists were not found to be statistically significant.

Clinically, as low-dose CT scans are non-contrast scans, the classically described contrast enhancement characteristics for diagnosing malignant nodules cannot be used to assess the risk of malignancy in these cases.

The availability of a highly sensitive nodule characterization tool will improve the early cancer detection rates. Radiologists aided by deep learning solutions for malignancy have the potential to identify lung cancer earlier as well as reduce unnecessary biopsies.

Read full poster

Synthetic PET Generator: A Novel Method to Improve Lung Nodule Detection by Combining Outputs from a Pix2pix Conditional Adversarial Network and a Convolutional Neural Network-Based Malignancy Probability Estimator

In collaboration with Centre for Advanced Research in Imaging, Neurosciences and Genomics, New Delhi, India; 

ECRPresented: 27 November 2018, RSNA 2018

poster

Background

Assessment of malignancy of lung nodules on CT scans is a subjective and arduous task for radiologists with low reported accuracy rates, especially for small nodules. We propose a novel method to generate a synthetic PET image of the lung from CT images using Conditional Generative Adversarial Networks (cGAN) that can improve the sensitivity of the radiologist in the detection of malignant lung nodules.

Evaluation

We used a combination of a PET Generator combined with a Malignancy Probability Estimator to generate a synthetic PET image from Lung CT scan. The PET Generator is a conditional adversarial network (pix2pix) trained on slices containing the Lung from 100 PET-CT scans which were acquired on patients suspected or diagnosed with Lung Cancer. The model performed at a mean squared error of 0.08 when compared in SUV units. The malignancy probability estimator is a 20-layer deep residual convolutional neural network trained on a dataset of 1595 scans from the NLST trial. The model performed produced a ROC of 0.89 when tested on 822 patients. The outputs of the PET Generator provides a background for overlaying outputs of the Malignancy Probability Estimator which together produce the synthetic PET image.

Discussion

When tested on a dataset of 30 images, the synthetic PET model performed at a mean squared error of 0.08 when compared in SUV units. The malignancy model was independently tested on 350 scans and produced an AUC of 0.89. A dataset of 22 malignant scans is used to benchmark performance of malignancy detection. Using the CT scan alone, three radiologists had sensitivities of 86%, 81% and 72% in detecting malignant studies. Using the synthetic PET as an additional modality, an increased sensitivity of 95% can be obtained. However, it is important to note that the SUV values detected on the nodules were not correlated with the actual SUV values.

Conclusion

Synthetic PET images can potentially increase the sensitivity of malignant nodule detection from Lung CT images. Such a modality can be easier for radiologists to understand than naive probability heatmaps. Further research is required to investigate the effect of potential biases and investigate appropriate clinical application.

Longitudinal Multiple Sclerosis Lesion Segmentation: Resource & Challenge

Contributed as Team IIT Madras, representing Indian Institute of Technology, Madras, India; 

NeuroimagePublished: 11 January 2017, Neuroimage

Journal

Rationale and Objectives

To explain predictions of a deep residual convolutional network for characterization of lung nodule by analyzing heat maps.

Materials and Methods

A 20-layer deep residual CNN was trained on 1245 Chest CTs from National Lung Screening Trial (NLST) trial to predict the malignancy risk of a nodule. We used occlusion to systematically block regions of a nodule and map drops in malignancy risk score to generate clinical attribution heatmaps on 103 nodules from Lung Image Database Consortium image collection and Image Database Resource Initiative (LIDC-IDRI) dataset, which were analyzed by a thoracic radiologist. The features were described as heat inside nodule -bright areas inside nodule, peripheral heat continuous/interrupted bright areas along nodule contours, heat in adjacent plane -brightness in scan planes juxtaposed with the nodule, satellite heat - a smaller bright spot in proximity to nodule in the same scan plane, heat map larger than nodule bright areas corresponding to the shape of the nodule seen outside the nodule margins and heat in calcification.

Results

These six features were assigned binary values. This feature vector was fedinto a standard J48 decision tree with 10-fold cross-validation, which gave an 85 % weighted classification accuracy with a 77.8% True Positive (TP) rate, 8% False Positive (FP) rate for benign cases and 91.8% TP and 22.2% FP rates for malignant cases. Heat Inside nodule was more frequently observed in nodules classified as malignant whereas peripheral heat, heat in adjacent plane, and satellite heat were more commonly seen in nodules classified as benign.

Conclusion

We discuss the potential ability of a radiologist to visually parse the deep learning algorithm generated "heat map" to identify features aiding classification.

Read full paper

Sign up for early access