Benchmarking Eliminative Radiomic Feature Selection for Head and Neck Lymph Node Classification

Abstract

In head and neck squamous cell carcinoma (HNSCC) pathologic cervical lymph nodes (LN) remain important negative predictors. Current criteria for LN-classification in contrast-enhanced computed-tomography scans (contrast-CT) are shape-based; contrast-CT imagery allows extraction of additional quantitative data ("features"). The data-driven technique to extract, process, and analyze features from contrast-CTs is termed "radiomics". Extracted features from contrast-CTs at various levels are typically redundant and correlated. Current sets of features for LN-classification are too complex for clinical application. Effective eliminative feature selection (EFS) is a crucial preprocessing step to reduce the complexity of sets identified. We aimed at exploring EFS-algorithms for their potential to identify sets of features, which were as small as feasible and yet retained as much accuracy as possible for LN-classification. In this retrospective cohort-study, which adhered to the STROBE guidelines, in total 252 LNs were classified as "non-pathologic" (n = 70), "pathologic" (n = 182) or "pathologic with extracapsular spread" (n = 52) by two experienced head-and-neck radiologists based on established criteria which served as a reference. The combination of sparse discriminant analysis and genetic optimization retained up to 90% of the classification accuracy with only 10% of the original numbers of features. From a clinical perspective, the selected features appeared plausible and potentially capable of correctly classifying LNs. Both the identified EFS-algorithm and the identified features need further exploration to assess their potential to prospectively classify LNs in HNSCC.

Keywords: computed-tomography; extracapsular spread; feature extraction; genetic algorithms; head and neck squamous carcinoma; lymph nodes; radiomics; recursive feature elimination; sparse discriminant analysis.

Conflict of interest statement

The authors declare no conflict of interest.

Figures

Figure 1

Study flow and patient inclusion modified according to STARD criteria [19]. A total of 1110 patients were potentially eligible, of which 823 did not meet the inclusion criteria. Of 287 eligible patients a representative random sample of 28 patients (~10%) was drawn. The clinical characteristics of the 28 included patients are presented in Table 1.

Figure 2

Example of a LN classified as “pathologic” in a staging-CT of a 55-year-old, male HNSCC patient with a tumor of the oral cavity staged cT4a cN2b cM0. Original segmentation in the axial plane (A); sagittal (B) and coronal (C) reformatted views and three-dimensional rendering (D) of the LN are provided by the software. Dashed arrows show central necrosis; no soft tissue infiltration and irregular LN capsule was observed.

Figure 3

Example of a LN classified as “pathologic with ECS” in a Staging-CT of a 53-year-old, male HNSCC-patient of the oral cavity staged cT2 cN3b cM0. Manual segmentation in the axial plane (A); sagittal (B), coronal (C) reformatted views and three-dimensional rendering (D) of the LN are provided by the software. Solid arrows show soft tissue infiltration and an irregular LN capsule. Dashed arrows show central necrosis.

Figure 4

Example of a LN classified as “non-pathologic” in a staging-CT of a 65-year-old, female HNSCC-patient with a tumor of the oral cavity staged cT4a cN2c cM0. Manually segmented LN in the axial plane (A); sagittal (B), coronal (C) reformatted views and three-dimensional rendering (D) of the LN are provided by the software. Neither central necrosis nor soft tissue infiltration nor irregular LN capsule were observed for this LN.

Figure 5

Balanced accuracy distributions of EFS-algorithms trained and evaluated on the LN-label “pathologic”. For reference, the BACC of the LDA classifier without feature count reduction is shown as a dash-dotted green horizontal line (at a value of 0.87). The combination of RFE and GA (black box) appeared to be the potentially most useful EFS-algorithms, retaining a diagnostic accuracy of >80% with only 10% of the features.

Figure 6

Balanced accuracy distributions of EFS-algorithms trained and evaluated on the LN-label “pathologic with ECS”. For reference, the BACC of the LDA classifier without feature count reduction is shown as a dash-dotted green horizontal line (at a value of 0.96). The combination of SDA and GA (black box) appeared to be the potentially most useful EFS-algorithms, retaining a diagnostic accuracy of >90% with only 10% of the features.

Figure 7

Balanced accuracy distributions of EFS-algorithms trained and evaluated on the LN-label “non-pathologic”. For reference, the BACC of the LDA classifier without feature count reduction (dash-dotted green horizontal line at 0.90). The combination of SDA and GA (black box) appeared to be the potentially most useful EFS-algorithms, retaining a diagnostic accuracy of >83% with only 10% of the features.