Deep learning enabled, deep ultraviolet scanning microscopy for intraoperative assessment of margin status during breast cancer surgery

点击翻译按钮获取中文摘要

Women with positive margins after breast-conserving surgery (BCS) have a 2-fold increased risk of cancer recurrence and are recommended to undergo additional re-excision surgery to achieve negative margins. Additional surgery is associated with significant emotional, cosmetic and financial burdens for patients and their caregivers. Although radiography, frozen section, touch prep and MarginProbe are available for intraoperative margin assessment, their accuracy is variable and most, except radiographic examination, are time- and labor- intensive and not routinely used. Since publication of the 2014 SSO-ASTRO guidelines for invasive cancer recommending re-excision for positive margins only, the re-excision rates have decreased but remain substantial (14-18%) with significant variation among surgeons. Because the size of BCS specimens varies significantly (a few to >40 cm2 per margin) and positive margins often include one or multiple sites/foci, a device with both variable margin coverage and microscopic resolution that can accurately evaluate an entire surgical specimen within a few minutes is highly desirable. While new technologies have been proposed, they are either point or high resolution devices with a very small field-of-view that requires excessive time to scan a specimen, or wide- field devices with low resolution and poor sensitivity. None has demonstrated the capability of analyzing an entire lumpectomy specimen with both adequate resolution and time efficiency in a clinical setting. Our goal is to develop a deep learning (DL) enabled, deep ultraviolet (DUV) scanning microscope (DDSM) for subcellular resolution and rapid (<5 min) examination of freshly excised tumor specimens during BCS. We hypothesize that there are significant subcellular optical contrasts that can be identified by the DDSM to differentiate breast cancer cells from normal tissue. Our preliminary DUV images demonstrate excellent contrasts and accuracy for identification of breast cancer cells. We propose that large and variable margin coverage, microscopic resolution and high speed are achieved by using: 1) DUV light for surface excitation of fresh specimens; 2) parallel imaging of two margins; 3) a low optical manification for fast speed; and 4) DL and sparse-sampling (SS) to rapidly search for pathological features of cancer cells. In Aim 1, a novel DDSM instrument will be developed and used to image 120 fresh breast tissues. DL classification algorithms will be developed and validated using the 120 tissue samples in Aim 2. Aim 3 will integrate DL and SS algorithms into the DDSM and demonstrate for fast detection of variable amount of cancer cells on the surfaces of breast tumor specimens. DDSM is highly innovative, combining DUV microscopy, parallel imaging, DL classification, and SS in a fast, compact, automated design. During initial BCS, if the DDSM accurately and efficiently identifies positive margins, additional breast tissue would be removed from the surgical cavity until negative margins are achieved and unnecessary removal of additional tissue would be avoided, thus decrease the need for additional surgery. DDSM is a platform technology that can be used with other imaging modalities or adapted for detection of other cancer or noncancer conditions.