Cytomorphological criteria for separation of pulmonary adenocarcinomas from squamous cell carcinomas: A statistical learning approach

Background Current therapy requires separation of non-small cell carcinomas into adenocarcinomas (AC) and squamous cell carcinomas (SCC). A meta-analysis has shown a pooled diagnostic sensitivity of 63% and specificity of 95% for the diagnosis of AC. While a number of cytomorphological features have been proposed for separation of AC from SCC, we are unaware of a statistically based analysis of cytomorphological features useful for separation of these two carcinomas. We performed logistic regression analysis of cytological features useful in classifying SCC and AC. Design Sixty-one Papanicolaou-stained fine needle aspiration specimens (29 AC/32 SCC) were reviewed by two board-certified cytopathologists for nine features (eccentric nucleoli, vesicular chromatin, prominent nucleoli, vacuolated cytoplasm, 3-dimensional cell balls, dark non-transparent chromatin, central nucleoli, single malignant cells and spindle-shaped cells). All cytological specimens had surgical biopsy results. Inter-rater agreement was assessed by Cohen's kappa. Association between features and AC was determined using hierarchical logistic regression model where feature scores were nested within reviewers. A model to classify cases as SCC or AC was developed and verified by k-fold verification (k = 5). Classification performance was assessed using the area under the receiver operating characteristic curve. Results Observed rater agreement for scored features ranged from 49% to 82%. Kappa scores were clustered in three groups. Raters demonstrated good agreement for prominent nucleoli, vesicular chromatin and eccentric nuclei. Fair agreement was seen for 3-dimensional cell balls, dark non-transparent chromatin, and presence of spindle-shaped cells. Association of features with adenocarcinoma showed four statistically significant associations (P < 0.001) with adenocarcinoma. These features were prominent nucleoli, vesicular chromatin, eccentric nuclei and three-dimensional cell balls. Spindle-shaped cells and dark non-transparent chromatin were negatively associated with adenocarcinoma. Conclusions Logistic regression analysis demonstrated six features helpful in separation of AC from SCC. Prominent nucleoli, vesicular chromatin, cell balls and eccentric nucleoli were positively associated with AC and demonstrated a P value of 0.001 or less. The presence of dark, non-transparent chromatin and spindle-shaped cells favoured the diagnosis of SCC.