309P - Improved diagnostic accuracy in MRI breast lesions using a classification system and multilayer perceptron neural network

Background

The American College of radiology proposed BI-RADS lexicon lacks defined rules which direct conversion of specific imaging features into a diagnostic category, results in a discrepancy of reporting. This study compares results from multilayer Perceptron neural network and a classification tree.

Methods

A total of 316 lesions with successive histological verification (221 malignant, 95 benign) were investigated. Six lesion criteria's were assessed by 2 readers in consensus. Classification analysis was performed using the chi-squared automatic interaction detection (CHAID) method. Results include the probability for malignancy for every descriptor combination in the classification tree. Simultaneously a multilayer Perceptron neural network was developed by using SPSS software.

Results

A classification tree incorporating 6 lesion descriptors with a depth of 4 ramifications (1- ADC values; 2 -root sign; 3- enhancement pattern; 4 - oedema) was calculated. Of all 316 lesions, 38 (40 %) and 212 (95.9 %) could be classified as benign and malignant with an accuracy above 95 %, respectively. Overall diagnostic accuracy was 79.1 %. The multilayer perceptron network segregated the lesions into training and testing sets in a ratio of 7:3. With a hyperbolic tangent activation function, there were six units of hidden layer and the model show a 20% and 17% incorrect predictions in the training in the testing sets. The diagnostic accuracy of malignant and benign lesions was 92% and 52 % in both the training and testing sets. The area under the curve of the ROC was .855. The order of importance of synaptic weights calculated from the model were ADC ( 0.257), Internal enhancement (0.233), ROOT SIGN (0.175), Margins (0.138), Curve type (0.138), edema (0.038) and mass / non mass (0.021).

Conclusions

The classification algorithm correctly classified 95 % malignant lesions with accuracy above 95 %. The neural network model showed good results on internal validation and revealed ADC to be the most significant parameter with the least importance to morphological classification into mass and non-mass lesions. Also, the dynamic contrast curve patterns were more significant than margins.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Rajiv Gandhi Cancer Institute and Research Center.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.