Artificial intelligence methods including deep neural networks (DNN) can provide rapid molecular classification of tumors from routine histology with accuracy that matches or exceeds human pathologists. Discerning how neural networks make their predictions remains a significant challenge, but explainability tools help provide insights into what models have learned when corresponding histologic features are poorly defined. Here, we present a method for improving explainability of DNN models using synthetic histology generated by a conditional generative adversarial network (cGAN). We show that cGANs generate high-quality synthetic histology images that can be leveraged for explaining DNN models trained to classify molecularly-subtyped tumors, exposing histologic features associated with molecular state. Fine-tuning synthetic histology through class and layer blending illustrates nuanced morphologic differences between tumor subtypes. Finally, we demonstrate the use of synthetic histology for augmenting pathologist-in-training education, showing that these intuitive visualizations can reinforce and improve understanding of histologic manifestations of tumor biology.

No abstract

6007 Background: The role of neoadjuvant immunotherapy in curative-intent head and neck squamous cell carcinoma (HNSCC) remains poorly defined. Survival for locoregionally advanced (LA) HPV negative (-) HNSCC remains poor with two-year survival of ~50%, and substantial treatment-related toxicity with standard chemoradiation (CRT). Given the activity of anti-PD1 in recurrent/metastatic HNSCC, we studied neoadjuvant nivolumab with chemotherapy and the feasibility of subsequent response-stratified CRT in HPV(-) LA HNSCC. Methods: The DEPEND trial (NCT03944915) is a phase II trial of nivolumab, paclitaxel, and carboplatin followed by response-stratified CRT for previously untreated stage IVA-B (AJCC-8th edition) HPV(-) HNSCC. The ultimate goal is to evaluate radiation volume and/or dose reduction to decrease long-term toxicities. Eligible patients received three 21-day cycles of nivolumab 360mg day 1, paclitaxel 100mg/m2 on days 1/8/15, and carboplatin AUC5 day 1. Patients with ≥50% reduction by RECIST 1.1 received response-adapted CRT to 66Gy with elimination of elective nodal volumes; < 50% reduction received standard-dose CRT to 70-75Gy. Post-CRT nivolumab 480mg every 4 weeks for 9 months was administered. The primary endpoint was deep response rate (DRR) defined as the proportion of patients with ≥50% reduction. Tumor PD-L1 immunohistochemistry was reported as combined positive score (CPS). Results: Thirty-six eligible patients started treatment between September 2019 and June 2022. Median age 59 (range 27-77), 22% female, 80% 20PYH smoking, 39% oral cavity, 19% oropharynx, 25% larynx/hypopharynx, 78% T3/4 and 78% N2/3. PD-L1 CPS ≥1 in 58%. The DRR with nivolumab/chemotherapy was 54% (95% CI 0.37-0.72), which met our statistical endpoint. The ORR was 89%. CRT stratification was as follows: Response-adapted CRT (n = 19) and standard-dose CRT (n = 16). At a median follow-up of 14 months, 2-year PFS and OS were 64% (95%CI 0.41-0.80) and 76% (95%CI 0.53-0.89), respectively. By CRT stratification, 2-year PFS was 79% and 46% in response-adapted and standard-dose CRT, and 2-year OS was 86% and 67% in response-adapted and standard-dose CRT, respectively. One patient died from disease progression during neoadjuvant therapy. 2-year distant control in response-adapted and standard-dose CRT arms was 100% and 63%, and 2-year locoregional control was 85% and 92%, respectively. PD-L1 CPS ≥1 and < 1 demonstrated DRR of 75% and 27%, respectively ( p= 0.01). Conclusions: Nivolumab-based neoadjuvant chemoimmunotherapy led to deep responses, and response-adapted CRT was associated with favorable survival and locoregional control. PD-L1 expression was predictive of deep response to nivolumab-based neoadjuvant therapy. Late toxicity analysis between treatment arms is planned. Clinical trial information: NCT03944915 .