286P - Improved diagnostic accuracy on MR imaging in post-surgical recurrent head and neck SCC lesions using decision tree classification system

Background

Recurrent tumors on MRI described as intermediate T2 signal intensity and ADC quantitative values derived from diffusion sequences are prone to discrepancy. In post surgical setting of oral cavity ADC values are prone to artifacts limiting its utility. This investigation aims to build a decision model using quantitative robust parameters derived from MR imaging.

Methods

Four lesion quantitative parameters ( quantitative T2 lesion signal, T2 lesion/muscle signal ratio, T2 lesion /Tongue signal ratio and ADC values) were assessed in 68 lesions (54 malignant,14 benign). Classification analysis was performed using L1 regularization of features in a Logistic regression, Statistical feature selection methods like ANOVA f-value and chi square and lastly a Entropy based feature selection using decision tree. Results include the probability for malignancy for every descriptor combination in the classification tree. Area under the curve (AUC) used as the statistical parameters to find model efficiency was calculated using software "R".

Results

Logistic regression based classifier could predict the probability of cancer based on T2 based features alone. ADC was not found helpful in predicting the disease. Both scores obtained from ANOVA and Chi-square have a different assumptions about distributions of input feature values and output probabilities, but yielded different scores. Both methods point to T2 as most significant in predicting output probabilities of cancer. Lastly, the decision tree showed T2 based features in addition to ADC provide maximum diagnostic value in determining cancer in patients. The area under the curve of the ROC was .940 for additive T2 and ADC and only 0.74 for ADC values alone. The signal ratios (T2 lesion/muscle signal ratio and T2 lesion /Tongue signal ratio) have an AUC 0.96.

Conclusions

Though each method of feature selection has certain shortfalls due to the assumptions but results demonstrate T2 feature outranking all others, indicating its high predictive power in determining the probability of disease. It is therefore possible to train predictive robust models based on T2 quantitative features with high level of accuracy and precision.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.