Basal Cell Carcinoma of Skin, Squamous Cell Carcinoma (Skin)
Conditions
Keywords
two-photon microscopy, machine learning
Brief summary
The goal of this study is to investigate the ability of a machine learning model to evaluate two-photon fluorescence microscopy images of dermatologic biopsies at point of care. The main question it aims to answer is: • How well do two-photon fluorescence images of biopsies taken in a clinic and evaluated by a machine learning model agree with conventional histology?
Detailed description
This study will image biopsy specimens at point of care using two-photon fluorescence microscopy (TPFM) and then assess how well the images predict the eventual clinical diagnosis using a machine learning model. Because two-photon images can be acquired from small biopsy specimens within minutes of excision, they could potentially be used to immediately diagnose patients, but the accuracy of TPFM for various skin conditions is unknown. Individual biopsy specimens in a dermatology clinic will be imaged using TPFM shortly after biopsy procedures. Immediately following imaging, a machine learning model will evaluate the TPFM images then compute a confidence score for a diagnosis of basal cell carcinoma (BCC), squamous cell carcinoma, and non-cancer. The relative confidence in each diagnosis will be compared, and if sufficient confidence is achieved, the model will render a diagnosis or else flag the specimen as indeterminate for manual pathologist review. This workflow will evaluate the use of ML + TPFM to perform point of care diagnosis of skin lesions. Following TPFM imaging, the specimen will be submitted for histological processing, which will guide actual patient treatment. Following conclusion of patient treatment, the resulting histology slides will be scanned for comparison and the final patient diagnosis recorded. Images of the histology slides will be read by a pathologist to establish a gold-standard diagnosis. The official diagnosis and the diagnosis from the collaborating pathologist will be compared. Patient treatment will still be decided by conventional histopathology. TPFM will not be used to change treatment.
Interventions
Ex vivo tissues will be imaged with two-photon microscopy and analyzed with machine learning for diagnosis
Sponsors
University of Rochester
National Cancer Institute (NCI)
Rochester Dermatologic Surgery
Study design
Allocation
NA
Intervention model
SINGLE_GROUP
Primary purpose
DIAGNOSTIC
Masking
NONE
Eligibility
Sex/Gender
ALL
Healthy volunteers
No
Inclusion criteria
* Punch, excisional or shave biopsy specimen
Exclusion criteria
* Biopsy indication includes melanoma or dysplastic/atypical nevus * Excision thickness of less than 1 mm * Excision longest dimension less than 2 mm * Excision performed as multiple pieces in a single specimen container
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Sensitivity of Machine Learning Analysis of Two Photon Fluorescence Microscopy Images At Point of Care | During or immediately following patient biopsy (same day) | A machine learning model will evaluate TPFM images of patient biopsies at point of care. Sensitivity will be calculated for the machine learning model using two photon fluorescence microscopy images. Sensitivity is defined as the number of true positive diagnoses divided by the sum of true positive and false negative diagnoses among biopsy specimens for which the machine learning model provides a definitive diagnosis. The patient's ultimate clinical diagnosis will serve as the reference standard. |
| Specificity of Machine Learning Analysis of Two Photon Fluorescence Microscopy Images At Point of Care | During or immediately following patient biopsy (same day) | A machine learning model will evaluate TPFM images of patient biopsies at point of care. Specificity will be calculated for the machine learning model using two photon fluorescence microscopy images. Specificity is defined as the number of true negative diagnoses divided by the sum of true negative and false positive diagnoses among biopsy specimens for which the machine learning model provides a definitive diagnosis. The patient's ultimate clinical diagnosis will serve as the reference standard. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Proportion of Discordant Diagnoses Attributable to Machine Learning Model Interpretation Errors | After completion of patient diagnosis (typically 1-2 weeks after procedure) | For biopsy specimens with discordant diagnoses between the machine learning model and the patient's ultimate clinical diagnosis, a dermatopathologist will review each case and classify the source of disagreement as machine learning model interpretation error, image quality limitation, or image coregistration error. The proportion of discordant diagnoses attributable to each source of disagreement will be reported. |
| Proportion of Biopsy Specimens With a Definitive Machine Learning Diagnosis | During or immediately following patient biopsy (same day) | The proportion of biopsy specimens for which the machine learning model provides a definitive diagnosis based on two photon fluorescence microscopy images will be calculated as the number of specimens receiving a definitive diagnosis divided by the total number of specimens evaluated. |
Countries
United States
Contacts
CONTACTMichael Giacomelli, Ph.D
Outcome results
None listed