Predicting Non-Small-Cell Lung Cancer Survival after Curative Surgery via Deep Learning of Diffusion MRI

BACKGROUND

the objective of this study is to evaluate the predictive power of the survival model using deep learning of diffusion-weighted images (DWI) in patients with non-small-cell lung cancer (NSCLC).

METHODS

DWI at b-values of 0, 100, and 700 sec/mm2 (DWI0, DWI100, DWI700) were preoperatively obtained for 100 NSCLC patients who underwent curative surgery (57 men, 43 women; mean age, 62 years). The ADC0-100 (perfusion-sensitive ADC), ADC100-700 (perfusion-insensitive ADC), ADC0-100-700, and demographic features were collected as input data and 5-year survival was collected as output data. Our survival model adopted transfer learning from a pre-trained VGG-16 network, whereby the softmax layer was replaced with the binary classification layer for the prediction of 5-year survival. Three channels of input data were selected in combination out of DWIs and ADC images and their accuracies and AUCs were compared for the best performance during 10-fold cross validation.

RESULTS

66 patients survived, and 34 patients died. The predictive performance was the best in the following combination: DWI0-ADC0-100-ADC0-100-700 (accuracy: 92%; AUC: 0.904). This was followed by DWI0-DWI700-ADC0-100-700, DWI0-DWI100-DWI700, and DWI0-DWI0-DWI0 (accuracy: 91%, 81%, 76%; AUC: 0.889, 0.763, 0.711, respectively). Survival prediction models trained with ADC performed significantly better than the one trained with DWI only (p-values < 0.05). The survival prediction was improved when demographic features were added to the model with only DWIs, but the benefit of clinical information was not prominent when added to the best performing model using both DWI and ADC.

CONCLUSIONS

Deep learning may play a role in the survival prediction of lung cancer. The performance of learning can be enhanced by inputting precedented, proven functional parameters of the ADC instead of the original data of DWIs only.

Radiomics analysis based on multiple parameters MR imaging in the spine: Predicting treatment response of osteolytic bone metastases to chemotherapy in breast cancer patients

PURPOSE

To assess the value of radiomics, apparent diffusion coefficient (ADC), intravoxel incoherent motion (IVIM) and stretched-exponential (SE) MR imaging in prediction of therapeutic response in patients with spinal metastases before chemotherapy.

METHODS

Thirty-six patients with 190 osteolytic metastatic lesions from breast cancer were prospectively enrolled and underwent MR imaging before and after 6 months' treatment on a 1.5 T MRI. According to MDA criteria, 68 lesions were categorized as progressive disease (PD) and 122 lesions were categorized as stable or improvement (non-PD). The regions of interest (ROIs) were manually drawn on DWI, T1WI, T2WI and FS-T2WI by two radiologists with ITK-SNAP. The ADC_all (multiple b-values method), IVIM parameters (D, D* and f) and SE parameters (DDC and α) were generated. The radiomics features were selected from the ROIs.

RESULTS

The mean values of ADC, DDC, and D before treatment were significantly higher in non-PD group than those in PD group (P = 0.001). The radiomics based on ADC_all had the highest AUC value (0.852), followed by that of the T2WI (0.829) and FS-T2WI (0.798). The radiomics model using ADC_all and FS-T2WI showed excellent efficiency in predicting treatment response with AUCs of 0.905 and 0.873 in training and validation cohorts. The radiomics model had better performance than that of ADC_all, D, and DDC for predicting treatment response of bone metastases.

CONCLUSION

Radiomics model based on ADC_all and FS-T2WI could predict the treatment response and contribute to assisting clinicians in accurately choosing appropriated management.

Pseudoprogression in Glioblastoma: Role of Metabolic and Functional MRI-Systematic Review

Background: Glioblastoma is the most frequent malignant primitive brain tumor in adults. The treatment includes surgery, radiotherapy, and chemotherapy. During follow-up, combined chemoradiotherapy can induce treatment-related changes mimicking tumor progression on medical imaging, such as pseudoprogression (PsP). Differentiating PsP from true progression (TP) remains a challenge for radiologists and oncologists, who need to promptly start a second-line treatment in the case of TP. Advanced magnetic resonance imaging (MRI) techniques such as diffusion-weighted imaging, perfusion MRI, and proton magnetic resonance spectroscopic imaging are more efficient than conventional MRI in differentiating PsP from TP. None of these techniques are fully effective, but current advances in computer science and the advent of artificial intelligence are opening up new possibilities in the imaging field with radiomics (i.e., extraction of a large number of quantitative MRI features describing tumor density, texture, and geometry). These features are used to build predictive models for diagnosis, prognosis, and therapeutic response. Method: Out of 7350 records for MR spectroscopy, GBM, glioma, recurrence, diffusion, perfusion, pseudoprogression, radiomics, and advanced imaging, we screened 574 papers. A total of 228 were eligible, and we analyzed 72 of them, in order to establish the role of each imaging modality and the usefulness and limitations of radiomics analysis.

Histogram analysis in prostate cancer: a comparison of diffusion kurtosis imaging model versus monoexponential model

BACKGROUND

There is still little research about histogram analysis of diffusion kurtosis imaging (DKI) using in prostate cancer at present.

PURPOSE

To verify the utility of histogram analysis of DKI model in detection and assessment of aggressiveness of prostate cancer, compared with monoexponential model (MEM).

MATERIAL AND METHODS

Twenty-three patients were enrolled in this study. For DKI model and MEM, the D_app, K_app, and apparent diffusion coefficient (ADC) were obtained by using single-shot echo-planar imaging sequence. The pathologies were confirmed by in-bore magnetic resonance (MR)-guided biopsy. Regions of interest (ROI) were drawn manually in the position where biopsy needle was put. The mean values and histogram parameters in cancer and noncancerous foci were compared using independent-samples T test. Receiver operating characteristic curves were used to investigate the diagnostic efficiency. Spearman's test was used to evaluate the correlation of parameters and Gleason scores.

RESULTS

The mean, 10th, 25th, 50th, 75th, and 90th percentiles of ADC and D_app were significantly lower in prostate cancer than non-cancerous foci (P < 0.001). The mean, 50th, 75th, and 90th percentiles of K_app were significantly higher in prostate cancer (P < 0.05). There was no significant difference between the AUCs of two models (0.971 vs. 0.963, P > 0.05). With the increasing Gleason scores, the 10th ADC decreased (ρ = -0.583, P = 0.018), but the 90th K_app increased (ρ = 0.642, P = 0.007).

CONCLUSION

Histogram analysis of DKI model is feasible for diagnosing and grading prostate cancer, but it has no significant advantage over MEM.

B-Value Optimization in the Estimation of Intravoxel Incoherent Motion Parameters in Patients with Cervical Cancer

Objective

This study aimed to find the optimal number of b-values for intravoxel incoherent motion (IVIM) imaging analysis, using simulated and in vivo data from cervical cancer patients.

Materials and Methods

Simulated data were generated using literature pooled means, which served as reference values for simulations. In vivo data from 100 treatment-naïve cervical cancer patients with IVIM imaging (13 b-values, scan time, 436 seconds) were retrospectively reviewed. A stepwise b-value fitting algorithm calculated optimal thresholds. Feed forward selection determined the optimal subsampled b-value distribution for biexponential IVIM fitting, and simplified IVIM modeling using monoexponential fitting was attempted. IVIM parameters computed using all b-values served as reference values for in vivo data.

Results

In simulations, parameters were accurately estimated with six b-values, or three b-values for simplified IVIM, respectively. In vivo data showed that the optimal threshold was 40 s/mm² for patients with squamous cell carcinoma and a subsampled acquisition of six b-values (scan time, 198 seconds) estimated parameters were not significantly different from reference parameters (individual parameter error rates of less than 5%). In patients with adenocarcinoma, the optimal threshold was 100 s/mm², but an optimal subsample could not be identified. Irrespective of the histological subtype, only three b-values were needed for simplified IVIM, but these parameters did not retain their discriminative ability.

Conclusion

Subsampling of six b-values halved the IVIM scan time without significant losses in accuracy and discriminative ability. Simplified IVIM is possible with only three b-values, at the risk of losing diagnostic information.

Application of diffusion-weighted magnetic resonance imaging in the diagnosis of odontogenic lesions: a systematic review

OBJECTIVES

This systematic literature review addresses the use of diffusion-weighted magnetic resonance imaging (DWI) and apparent diffusion coefficient (ADC) for the evaluation of benign maxillomandibular odontogenic lesions.

STUDY DESIGN

Databases were searched, and original research studies or case report manuscripts up to April 2019 were included, using the keyword "diffusion," combined with the keywords "maxillofacial pathology," "oral pathology," "odontogenic tumors," "dental tissue neoplasms," "odontogenic cysts," and the histologic denomination of benign odontogenic lesions, according to the World Health Organization classification. Only English language articles and studies pertaining to DWI were selected.

RESULTS

Fifteen investigations (11 original articles and 4 case reports) of distinct benign odontogenic lesions were included. Most studies did not include exclusively odontogenic lesions in their samples.

CONCLUSIONS

It is too early to reach a conclusion that DWI and ADC can provide useful information in the differentiation of the histologic type of some benign odontogenic lesions on the basis of available data in the literature.

Development of voxel-based optimization diffusion kurtosis imaging (DKI) and comparison with conventional DKI

The aims of this study were to implement voxel-based optimization diffusion kurtosis imaging (DKI) and evaluate the accuracy of the method for the analysis of diffusion imaging data in comparison with conventional DKI. Conventional DKI and voxel-based optimization DKI were tested on a phantom and a human in a 1.5 T whole-body scanner. The differences in the diffusion coefficient (D) and diffusion kurtosis (K) values were analyzed using the Mann-Whitney U test, and the Holm correction was applied to the statistical analyses. In the phantom study, the D value resulting from voxel-based optimization DKI was significantly lower than those from conventional DKI in water and agarose solutions at concentrations of 50 and 100 g/L (all p < 0.01). Moreover, the K value was significantly lower in the water and agarose solutions at concentrations of 50, 100, and 200 g/L (all p < 0.01). In the human study, the D value resulting from voxel-based optimization DKI was significantly lower than that of conventional DKI in both white matter (WM), and gray matter (GM) (all p < 0.01). Moreover, the K value was significantly lower in cerebrospinal fluid, WM, and GM (all p < 0.01). To correctly measure the DKI, the optimized b values for each voxel must be used. Voxel-based optimization DKI is a method that optimizes the b values for each voxel. It appears that voxel-based optimization DKI improves the accuracy of the K value for biological tissues.

Diffusion-weighted magnetic resonance imaging of the paranasal sinuses: A systematic review

OBJECTIVES

This was a systematic review of studies on the use of diffusion-weighted imaging (DWI) for paranasal sinus diseases. The applications of DWI were analyzed along with the main results, and conclusions were obtained by the investigators.

STUDY DESIGN

Databases were searched using the keyword "diffusion" combined with "sinonasal," "paranasal sinus," "maxillary sinus," "frontal sinus," "ethmoid sinus," and "sphenoid sinus," including only articles that were published from 2008 to 2018. Only original English language studies with sinonasal disease samples were selected.

RESULTS

Sixteen studies about various sinonasal diseases were included. The main objectives of most of the studies were related to the use of the apparent diffusion coefficient (ADC) in the differentiation of benign lesions and malignant neoplasms. We concluded that the ADC for malignant neoplasms is lower. Histologic features of samples evaluated in the studies were heterogeneous.

CONCLUSIONS

The ADC may improve the quality of the diagnostic hypothesis, particularly in differentiating benign and malignant diseases. Furthermore, the differences between certain types of lesions could be determined by using the ADC. However, further studies focusing on inflammatory diseases should be performed. Overall, DWI and the ADC are promising methods that can be incorporated into routine evaluations.

Current contribution of diffusion tensor imaging in the evaluation of diffuse axonal injury

ABSTRACT Traumatic brain injury (TBI) is the number one cause of death and morbidity among young adults. Moreover, survivors are frequently left with functional disabilities during the most productive years of their lives. One main aspect of TBI pathology is diffuse axonal injury, which is increasingly recognized due to its presence in 40% to 50% of all cases that require hospital admission. Diffuse axonal injury is defined as widespread axonal damage and is characterized by complete axotomy and secondary reactions due to overall axonopathy. These changes can be seen in neuroimaging studies as hemorrhagic focal areas and diffuse edema. However, the diffuse axonal injury findings are frequently under-recognized in conventional neuroimaging studies. In such scenarios, diffuse tensor imaging (DTI) plays an important role because it provides further information on white matter integrity that is not obtained with standard magnetic resonance imaging sequences. Extensive reviews concerning the physics of DTI and its use in the context of TBI patients have been published, but these issues are still hazy for many allied-health professionals. Herein, we aim to review the current contribution of diverse state-of-the-art DTI analytical methods to the understanding of diffuse axonal injury pathophysiology and prognosis, to serve as a quick reference for those interested in planning new studies and who are involved in the care of TBI victims. For this purpose, a comprehensive search in Pubmed was performed using the following keywords: “traumatic brain injury”, “diffuse axonal injury”, and “diffusion tensor imaging”.

[Effect of Slice Thickness for Apparent Diffusion Coefficient Measurement of Mass]

Apparent diffusion coefficient (ADC) values calculated from diffusion-weighted magnetic resonance imaging (DW-MRI) can be used for differentiation of tumors. Clinically, ADC values are used for monitoring treatment response after chemotherapy or radiation. However, it is reported that the threshold of the ADC value differs among institutions. In addition, there are reports regarding the change factor of the ADC value. Slice thickness may induce error in the ADC value by the influence of the partial volume effect in thicker objects, and by the influence of signal-to-noise ratio (SNR) in thinner objects. Therefore, in this study, the effect of slice thickness was examined. The signal body of spherical high-diffusion coefficients of 6, 7.9, and 9.3 mm in diameter was fixed in the low-circumference material of the diffusion coefficient. These phantoms were imaged using DW imaging (DWI) of 1, 1.5, 2, 2.5, 3, 4, 5, 6, 7, 8, 9, 10, 15, and 20 mm slice thickness using the multi-b values. In addition, different SNR were imaged by changing field-of-view and the number of additions. ADC was calculated by DWI of the different b values. As a result, slice thickness showed a peak at 50-65% of the diameter of the signal body. Furthermore, ADC values fluctuated in the slice thickness in front of the peak with a change in SNR. In conclusion, the ADC value was most accurate at a setting of 50-65% of slice thickness for the object diameter.